Note
You are reading the documentation for MMOCR 0.x, which will soon be deprecated by the end of 2022. We recommend you upgrade to MMOCR 1.0 to enjoy fruitful new features and better performance brought by OpenMMLab 2.0. Check out the maintenance plan, changelog, code and documentation of MMOCR 1.0 for more details.
Welcome to MMOCR’s documentation!¶
You can switch between English and Chinese in the lower-left corner of the layout.
Installation¶
Prerequisites¶
Linux | Windows | macOS
Python 3.7
PyTorch 1.6 or higher
torchvision 0.7.0
CUDA 10.1
NCCL 2
GCC 5.4.0 or higher
Environment Setup¶
Note
If you are experienced with PyTorch and have already installed it, just skip this part and jump to the next section. Otherwise, you can follow these steps for the preparation.
Step 0. Download and install Miniconda from the official website.
Step 1. Create a conda environment and activate it.
conda create --name openmmlab python=3.8 -y
conda activate openmmlab
Step 2. Install PyTorch following official instructions, e.g.
On GPU platforms:
conda install pytorch torchvision -c pytorch
On CPU platforms:
conda install pytorch torchvision cpuonly -c pytorch
Installation Steps¶
We recommend that users follow our best practices to install MMOCR. However, the whole process is highly customizable. See Customize Installation section for more information.
Best Practices¶
Step 0. Install MMCV using MIM.
pip install -U openmim
mim install mmcv-full
Step 1. Install MMDetection as a dependency.
pip install mmdet
Step 2. Install MMOCR.
Case A: If you wish to run and develop MMOCR directly, install it from source:
git clone https://github.com/open-mmlab/mmocr.git
cd mmocr
pip install -r requirements.txt
pip install -v -e .
# "-v" increases pip's verbosity.
# "-e" means installing the project in editable mode,
# That is, any local modifications on the code will take effect immediately.
Case B: If you use MMOCR as a dependency or third-party package, install it with pip:
pip install mmocr
Step 3. (Optional) If you wish to use any transform involving albumentations
(For example, Albu
in ABINet’s pipeline), install the dependency using the following command:
# If MMOCR is installed from source
pip install -r requirements/albu.txt
# If MMOCR is installed via pip
pip install albumentations>=1.1.0 --no-binary qudida,albumentations
Note
We recommend checking the environment after installing albumentations
to
ensure that opencv-python
and opencv-python-headless
are not installed together, otherwise it might cause unexpected issues. If that’s unfortunately the case, please uninstall opencv-python-headless
to make sure MMOCR’s visualization utilities can work.
Refer to ‘albumentations`’s official documentation for more details.
Verify the installation¶
We provide two options to verify the installation via inference demo, depending on your installation method. You should be able to see a pop-up image and the inference result upon successful verification.

# Inference result
[{'filename': 'demo_text_det', 'text': ['yther', 'doyt', 'nan', 'heraies', '188790', 'cadets', 'army', 'ipioneered', 'and', 'icottages', 'land', 'hall', 'sgardens', 'established', 'ithis', 'preformer', 'social', 'octavial', 'hill', 'pm', 'ct', 'lof', 'aborought']}]
Case A - Installed from Source¶
Run the following in MMOCR’s directory:
python mmocr/utils/ocr.py --det DB_r18 --recog CRNN demo/demo_text_det.jpg --imshow
Case B - Installed as a Package:¶
Step 1. We need to download configs, checkpoints and an image necessary for the verification.
mim download mmocr --config dbnet_r18_fpnc_1200e_icdar2015 --dest .
mim download mmocr --config crnn_academic_dataset --dest .
wget https://raw.githubusercontent.com/open-mmlab/mmocr/main/demo/demo_text_det.jpg
The downloading will take several seconds or more, depending on your network environment. The directory tree should look like the following once everything is done:
├── crnn_academic-a723a1c5.pth
├── crnn_academic_dataset.py
├── dbnet_r18_fpnc_1200e_icdar2015.py
├── dbnet_r18_fpnc_sbn_1200e_icdar2015_20210329-ba3ab597.pth
└── demo_text_det.jpg
Step 2. Run the following codes in your Python interpreter:
from mmocr.utils.ocr import MMOCR
ocr = MMOCR(recog='CRNN', recog_ckpt='crnn_academic-a723a1c5.pth', recog_config='crnn_academic_dataset.py', det='DB_r18', det_ckpt='dbnet_r18_fpnc_sbn_1200e_icdar2015_20210329-ba3ab597.pth', det_config='dbnet_r18_fpnc_1200e_icdar2015.py')
ocr.readtext('demo_text_det.jpg', imshow=True)
Customize Installation¶
CUDA versions¶
When installing PyTorch, you need to specify the version of CUDA. If you are not clear on which to choose, follow our recommendations:
For Ampere-based NVIDIA GPUs, such as GeForce 30 series and NVIDIA A100, CUDA 11 is a must.
For older NVIDIA GPUs, CUDA 11 is backward compatible, but CUDA 10.2 offers better compatibility and is more lightweight.
Please make sure the GPU driver satisfies the minimum version requirements. See this table for more information.
Note
Installing CUDA runtime libraries is enough if you follow our best practices, because no CUDA code will be compiled locally. However if you hope to compile MMCV from source or develop other CUDA operators, you need to install the complete CUDA toolkit from NVIDIA’s website, and its version should match the CUDA version of PyTorch. i.e., the specified version of cudatoolkit in conda install
command.
Install MMCV without MIM¶
MMCV contains C++ and CUDA extensions, thus depending on PyTorch in a complex way. MIM solves such dependencies automatically and makes the installation easier. However, it is not a must.
To install MMCV with pip instead of MIM, please follow MMCV installation guides. This requires manually specifying a find-url based on PyTorch version and its CUDA version.
For example, the following command install mmcv-full built for PyTorch 1.10.x and CUDA 11.3.
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10/index.html
Install on CPU-only platforms¶
MMOCR can be built for CPU-only environment. In CPU mode you can train (requires MMCV version >= 1.4.4), test or inference a model.
However, some functionalities are gone in this mode:
Deformable Convolution
Modulated Deformable Convolution
ROI pooling
SyncBatchNorm
If you try to train/test/inference a model containing above ops, an error will be raised. The following table lists affected algorithms.
Operator | Model |
---|---|
Deformable Convolution/Modulated Deformable Convolution | DBNet (r50dcnv2), DBNet++ (r50dcnv2), FCENet (r50dcnv2) |
SyncBatchNorm | PANet, PSENet |
Using MMOCR with Docker¶
We provide a Dockerfile to build an image.
# build an image with PyTorch 1.6, CUDA 10.1
docker build -t mmocr docker/
Run it with
docker run --gpus all --shm-size=8g -it -v {DATA_DIR}:/mmocr/data mmocr
Dependency on MMCV & MMDetection¶
MMOCR has different version requirements on MMCV and MMDetection at each release to guarantee the implementation correctness. Please refer to the table below and ensure the package versions fit the requirement.
MMOCR | MMCV | MMDetection |
---|---|---|
main | 1.3.8 \<= mmcv \< 1.8.0 | 2.21.0 \<= mmdet \<= 3.0.0 |
0.6.3 | 1.3.8 \<= mmcv \< 1.8.0 | 2.21.0 \<= mmdet \<= 3.0.0 |
0.6.1, 0.6.2 | 1.3.8 \<= mmcv \<= 1.7.0 | 2.21.0 \<= mmdet \<= 3.0.0 |
0.6.0 | 1.3.8 \<= mmcv \<= 1.6.0 | 2.21.0 \<= mmdet \<= 3.0.0 |
0.5.0 | 1.3.8 \<= mmcv \<= 1.5.0 | 2.14.0 \<= mmdet \<= 3.0.0 |
0.4.0, 0.4.1 | 1.3.8 \<= mmcv \<= 1.5.0 | 2.14.0 \<= mmdet \<= 2.20.0 |
0.3.0 | 1.3.8 \<= mmcv \<= 1.4.0 | 2.14.0 \<= mmdet \<= 2.20.0 |
0.2.1 | 1.3.8 \<= mmcv \<= 1.4.0 | 2.13.0 \<= mmdet \<= 2.20.0 |
0.2.0 | 1.3.4 \<= mmcv \<= 1.4.0 | 2.11.0 \<= mmdet \<= 2.13.0 |
0.1.0 | 1.2.6 \<= mmcv \<= 1.3.4 | 2.9.0 \<= mmdet \<= 2.11.0 |
Getting Started¶
In this guide we will show you some useful commands and familiarize you with MMOCR. We also provide a notebook that can help you get the most out of MMOCR.
Installation¶
Check out our installation guide for full steps.
Dataset Preparation¶
MMOCR supports numerous datasets which are classified by the type of their corresponding tasks. You may find their preparation steps in these sections: Detection Datasets, Recognition Datasets, KIE Datasets and NER Datasets.
Inference with Pretrained Models¶
You can perform end-to-end OCR on our demo image with one simple line of command:
python mmocr/utils/ocr.py demo/demo_text_ocr.jpg --print-result --imshow
Its detection result will be printed out and a new window will pop up with result visualization. More demo and full instructions can be found in Demo.
Training¶
Training with Toy Dataset¶
We provide a toy dataset under tests/data
on which you can get a sense of training before the academic dataset is prepared.
For example, to train a text recognition task with seg
method and toy dataset,
python tools/train.py configs/textrecog/seg/seg_r31_1by16_fpnocr_toy_dataset.py --work-dir seg
To train a text recognition task with sar
method and toy dataset,
python tools/train.py configs/textrecog/sar/sar_r31_parallel_decoder_toy_dataset.py --work-dir sar
Training with Academic Dataset¶
Once you have prepared required academic dataset following our instruction, the only last thing to check is if the model’s config points MMOCR to the correct dataset path. Suppose we want to train DBNet on ICDAR 2015, and part of configs/_base_/det_datasets/icdar2015.py
looks like the following:
dataset_type = 'IcdarDataset'
data_root = 'data/icdar2015'
train = dict(
type=dataset_type,
ann_file=f'{data_root}/instances_training.json',
img_prefix=f'{data_root}/imgs',
pipeline=None)
test = dict(
type=dataset_type,
ann_file=f'{data_root}/instances_test.json',
img_prefix=f'{data_root}/imgs',
pipeline=None)
train_list = [train]
test_list = [test]
You would need to check if data/icdar2015
is right. Then you can start training with the command:
python tools/train.py configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py --work-dir dbnet
You can find full training instructions, explanations and useful training configs in Training.
Testing¶
Suppose now you have finished the training of DBNet and the latest model has been saved in dbnet/latest.pth
. You can evaluate its performance on the test set using the hmean-iou
metric with the following command:
python tools/test.py configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py dbnet/latest.pth --eval hmean-iou
Evaluating any pretrained model accessible online is also allowed:
python tools/test.py configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py https://download.openmmlab.com/mmocr/textdet/dbnet/dbnet_r18_fpnc_sbn_1200e_icdar2015_20210329-ba3ab597.pth --eval hmean-iou
More instructions on testing are available in Testing.
Demo¶
We provide an easy-to-use API for the demo and application purpose in ocr.py script.
The API can be called through command line (CL) or by calling it from another python script. It exposes all the models in MMOCR to API as individual modules that can be called and chained together. Tesseract is integrated as a text detector and/or recognizer in the task pipeline.
Example 1: Text Detection¶

Instruction: Perform detection inference on an image with the TextSnake recognition model, export the result in a json file (default) and save the visualization file.
CL interface:
python mmocr/utils/ocr.py demo/demo_text_det.jpg --output demo/ --det TextSnake --recog None --export demo/
Python interface:
from mmocr.utils.ocr import MMOCR
# Load models into memory
ocr = MMOCR(det='TextSnake', recog=None)
# Inference
results = ocr.readtext('demo/demo_text_det.jpg', output='demo/', export='demo/')
Example 2: Text Recognition¶

Instruction: Perform batched recognition inference on a folder with hundreds of image with the CRNN_TPS recognition model and save the visualization results in another folder. Batch size is set to 10 to prevent out of memory CUDA runtime errors.
CL interface:
python mmocr/utils/ocr.py %INPUT_FOLDER_PATH% --det None --recog CRNN_TPS --batch-mode --single-batch-size 10 --output %OUPUT_FOLDER_PATH%
Python interface:
from mmocr.utils.ocr import MMOCR
# Load models into memory
ocr = MMOCR(det=None, recog='CRNN_TPS')
# Inference
results = ocr.readtext(%INPUT_FOLDER_PATH%, output = %OUTPUT_FOLDER_PATH%, batch_mode=True, single_batch_size = 10)
Example 3: Text Detection + Recognition¶

Instruction: Perform ocr (det + recog) inference on the demo/demo_text_det.jpg image with the PANet_IC15 (default) detection model and SAR (default) recognition model, print the result in the terminal and show the visualization.
CL interface:
python mmocr/utils/ocr.py demo/demo_text_ocr.jpg --print-result --imshow
Note
When calling the script from the command line, the script assumes configs are saved in the configs/
folder. User can customize the directory by specifying the value of config_dir
.
Python interface:
from mmocr.utils.ocr import MMOCR
# Load models into memory
ocr = MMOCR()
# Inference
results = ocr.readtext('demo/demo_text_ocr.jpg', print_result=True, imshow=True)
Example 4: Text Detection + Recognition + Key Information Extraction¶

Instruction: Perform end-to-end ocr (det + recog) inference first with PS_CTW detection model and SAR recognition model, then run KIE inference with SDMGR model on the ocr result and show the visualization.
CL interface:
python mmocr/utils/ocr.py demo/demo_kie.jpeg --det PS_CTW --recog SAR --kie SDMGR --print-result --imshow
Note
Note: When calling the script from the command line, the script assumes configs are saved in the configs/
folder. User can customize the directory by specifying the value of config_dir
.
Python interface:
from mmocr.utils.ocr import MMOCR
# Load models into memory
ocr = MMOCR(det='PS_CTW', recog='SAR', kie='SDMGR')
# Inference
results = ocr.readtext('demo/demo_kie.jpeg', print_result=True, imshow=True)
API Arguments¶
The API has an extensive list of arguments that you can use. The following tables are for the python interface.
MMOCR():
Arguments | Type | Default | Description |
---|---|---|---|
det |
see models | PANet_IC15 | Text detection algorithm |
recog |
see models | SAR | Text recognition algorithm |
kie [1] |
see models | None | Key information extraction algorithm |
config_dir |
str | configs/ | Path to the config directory where all the config files are located |
det_config |
str | None | Path to the custom config file of the selected det model |
det_ckpt |
str | None | Path to the custom checkpoint file of the selected det model |
recog_config |
str | None | Path to the custom config file of the selected recog model |
recog_ckpt |
str | None | Path to the custom checkpoint file of the selected recog model |
kie_config |
str | None | Path to the custom config file of the selected kie model |
kie_ckpt |
str | None | Path to the custom checkpoint file of the selected kie model |
device |
str | None | Device used for inference, accepting all allowed strings by torch.device . E.g., 'cuda:0' or 'cpu'. |
[1]: kie
is only effective when both text detection and recognition models are specified.
Note
User can use default pretrained models by specifying det
and/or recog
, which is equivalent to specifying their corresponding *_config
and *_ckpt
. However, manually specifying *_config
and *_ckpt
will always override values set by det
and/or recog
. Similar rules also apply to kie
, kie_config
and kie_ckpt
.
readtext()¶
Arguments | Type | Default | Description |
---|---|---|---|
img |
str/list/tuple/np.array | required | img, folder path, np array or list/tuple (with img paths or np arrays) |
output |
str | None | Folder where visualization results are output for each image |
batch_mode |
bool | False | Whether use batch mode for inference [1] |
det_batch_size |
int | 0 | Batch size for text detection (0 for max size) |
recog_batch_size |
int | 0 | Batch size for text recognition (0 for max size) |
single_batch_size |
int | 0 | Batch size for only detection or recognition |
export |
str | None | Folder where the results of each image are exported |
export_format |
str | json | Format of the exported result file(s) |
details |
bool | False | Whether include the text boxes coordinates and confidence values |
imshow |
bool | False | Whether to show the result visualization on screen |
print_result |
bool | False | Whether to show the result for each image |
merge |
bool | False | Whether to merge neighboring boxes [2] |
merge_xdist |
float | 20 | The maximum x-axis distance to merge boxes |
[1]: Make sure that the model is compatible with batch mode.
[2]: Only effective when the script is running in det + recog mode.
All arguments are the same for the cli, all you need to do is add 2 hyphens at the beginning of the argument and replace underscores by hyphens.
(Example: det_batch_size
becomes --det-batch-size
)
For bool type arguments, putting the argument in the command stores it as true.
(Example: python mmocr/utils/ocr.py demo/demo_text_det.jpg --batch_mode --print_result
means that batch_mode
and print_result
are set to True
)
Models¶
Text detection:
Name | Reference | batch_mode inference support |
---|---|---|
DB_r18 | link | No |
DB_r50 | link | No |
DBPP_r50 | link | No |
DRRG | link | No |
FCE_IC15 | link | No |
FCE_CTW_DCNv2 | link | No |
MaskRCNN_CTW | link | No |
MaskRCNN_IC15 | link | No |
MaskRCNN_IC17 | link | No |
PANet_CTW | link | Yes |
PANet_IC15 | link | Yes |
PS_CTW | link | No |
PS_IC15 | link | No |
Tesseract | link | No |
TextSnake | link | Yes |
Text recognition:
Name | Reference | batch_mode inference support |
---|---|---|
ABINet | link | Yes |
CRNN | link | No |
CRNN_TPS | link | Yes |
MASTER | link | Yes |
NRTR_1/16-1/8 | link | Yes |
NRTR_1/8-1/4 | link | Yes |
RobustScanner | link | Yes |
SAR | link | Yes |
SAR_CN * | link | Yes |
SATRN | link | Yes |
SATRN_sm | link | Yes |
SEG | link | No |
Tesseract | link | No |
Warning
SAR_CN is the only model that supports Chinese character recognition and it requires a Chinese dictionary. Please download the dictionary from here for a successful run.
Key information extraction:
Name | Reference | batch_mode support |
---|---|---|
SDMGR | link | Yes |
Additional info¶
To perform det + recog inference (end2end ocr), both the
det
andrecog
arguments must be defined.To perform only detection set the
recog
argument toNone
.To perform only recognition set the
det
argument toNone
.details
argument only works with end2end ocr.det_batch_size
andrecog_batch_size
arguments define the number of images you want to forward to the model at the same time. For maximum speed, set this to the highest number you can. The max batch size is limited by the model complexity and the GPU VRAM size.MMOCR calls Tesseract’s API via
tesserocr
If you have any suggestions for new features, feel free to open a thread or even PR :)
Training¶
Training on a Single GPU¶
You can use tools/train.py
to train a model on a single machine with a CPU and optionally a GPU.
Here is the full usage of the script:
python tools/train.py ${CONFIG_FILE} [ARGS]
Note
By default, MMOCR prefers GPU to CPU. If you want to train a model on CPU, please empty CUDA_VISIBLE_DEVICES
or set it to -1 to make GPU invisible to the program. Note that CPU training requires MMCV >= 1.4.4.
CUDA_VISIBLE_DEVICES= python tools/train.py ${CONFIG_FILE} [ARGS]
ARGS | Type | Description |
---|---|---|
--work-dir |
str | The target folder to save logs and checkpoints. Defaults to ./work_dirs . |
--load-from |
str | Path to the pre-trained model, which will be used to initialize the network parameters. |
--resume-from |
str | Resume training from a previously saved checkpoint, which will inherit the training epoch and optimizer parameters. |
--no-validate |
bool | Disable checkpoint evaluation during training. Defaults to False . |
--gpus |
int | Deprecated, please use --gpu-id. Numbers of gpus to use. Only applicable to non-distributed training. |
--gpu-ids |
int*N | Deprecated, please use --gpu-id. A list of GPU ids to use. Only applicable to non-distributed training. |
--gpu-id |
int | The GPU id to use. Only applicable to non-distributed training. |
--seed |
int | Random seed. |
--diff-seed |
bool | Whether or not set different seeds for different ranks. |
--deterministic |
bool | Whether to set deterministic options for CUDNN backend. |
--cfg-options |
str | Override some settings in the used config, the key-value pair in xxx=yyy format will be merged into the config file. If the value to be overwritten is a list, it should be of the form of either key="[a,b]" or key=a,b. The argument also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]". Note that the quotation marks are necessary and that no white space is allowed. |
--launcher |
'none', 'pytorch', 'slurm', 'mpi' | Options for job launcher. |
--local_rank |
int | Used for distributed training. |
--mc-config |
str | Memory cache config for image loading speed-up during training. |
Training on Multiple GPUs¶
MMOCR implements distributed training with MMDistributedDataParallel
. (Please refer to datasets.md to prepare your datasets)
[PORT={PORT}] ./tools/dist_train.sh ${CONFIG_FILE} ${WORK_DIR} ${GPU_NUM} [PY_ARGS]
Arguments | Type | Description |
---|---|---|
PORT |
int | The master port that will be used by the machine with rank 0. Defaults to 29500. Note: If you are launching multiple distributed training jobs on a single machine, you need to specify different ports for each job to avoid port conflicts. |
CONFIG_FILE |
str | The path to config. |
WORK_DIR |
str | The path to the working directory. |
GPU_NUM |
int | The number of GPUs to be used per node. Defaults to 8. |
PY_ARGS |
str | Arguments to be parsed by tools/train.py . |
Training on Multiple Machines¶
You can launch a task on multiple machines connected to the same network.
NNODES=${NNODES} NODE_RANK=${NODE_RANK} PORT=${MASTER_PORT} MASTER_ADDR=${MASTER_ADDR} ./tools/dist_train.sh ${CONFIG_FILE} ${WORK_DIR} ${GPU_NUM} [PY_ARGS]
Arguments | Type | Description |
---|---|---|
NNODES |
int | The number of nodes. |
NODE_RANK |
int | The rank of current node. |
PORT |
int | The master port that will be used by rank 0 node. Defaults to 29500. |
MASTER_ADDR |
str | The address of rank 0 node. Defaults to "127.0.0.1". |
CONFIG_FILE |
str | The path to config. |
WORK_DIR |
str | The path to the working directory. |
GPU_NUM |
int | The number of GPUs to be used per node. Defaults to 8. |
PY_ARGS |
str | Arguments to be parsed by tools/train.py . |
Note
MMOCR relies on torch.distributed package for distributed training. Find more information at PyTorch’s launch utility.
Say that you want to launch a job on two machines. On the first machine:
NNODES=2 NODE_RANK=0 PORT=${MASTER_PORT} MASTER_ADDR=${MASTER_ADDR} ./tools/dist_train.sh ${CONFIG_FILE} ${WORK_DIR} ${GPU_NUM} [PY_ARGS]
On the second machine:
NNODES=2 NODE_RANK=1 PORT=${MASTER_PORT} MASTER_ADDR=${MASTER_ADDR} ./tools/dist_train.sh ${CONFIG_FILE} ${WORK_DIR} ${GPU_NUM} [PY_ARGS]
Note
The speed of the network could be the bottleneck of training.
Training with Slurm¶
If you run MMOCR on a cluster managed with Slurm, you can use the script slurm_train.sh
.
[GPUS=${GPUS}] [GPUS_PER_NODE=${GPUS_PER_NODE}] [CPUS_PER_TASK=${CPUS_PER_TASK}] [SRUN_ARGS=${SRUN_ARGS}] ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} ${CONFIG_FILE} ${WORK_DIR} [PY_ARGS]
Arguments | Type | Description |
---|---|---|
GPUS |
int | The number of GPUs to be used by this task. Defaults to 8. |
GPUS_PER_NODE |
int | The number of GPUs to be allocated per node. Defaults to 8. |
CPUS_PER_TASK |
int | The number of CPUs to be allocated per task. Defaults to 5. |
SRUN_ARGS |
str | Arguments to be parsed by srun. Available options can be found here. |
PY_ARGS |
str | Arguments to be parsed by tools/train.py . |
Here is an example of using 8 GPUs to train a text detection model on the dev partition.
./tools/slurm_train.sh dev psenet-ic15 configs/textdet/psenet/psenet_r50_fpnf_sbn_1x_icdar2015.py /nfs/xxxx/psenet-ic15
Running Multiple Training Jobs on a Single Machine¶
If you are launching multiple training jobs on a single machine with Slurm, you may need to modify the port in configs to avoid communication conflicts.
For example, in config1.py
,
dist_params = dict(backend='nccl', port=29500)
In config2.py
,
dist_params = dict(backend='nccl', port=29501)
Then you can launch two jobs with config1.py
ang config2.py
.
CUDA_VISIBLE_DEVICES=0,1,2,3 GPUS=4 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config1.py ${WORK_DIR}
CUDA_VISIBLE_DEVICES=4,5,6,7 GPUS=4 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config2.py ${WORK_DIR}
Commonly Used Training Configs¶
Here we list some configs that are frequently used during training for quick reference.
total_epochs = 1200
data = dict(
# Note: User can configure general settings of train, val and test dataloader by specifying them here. However, their values can be overridden in dataloader's config.
samples_per_gpu=8, # Batch size per GPU
workers_per_gpu=4, # Number of workers to process data for each GPU
train_dataloader=dict(samples_per_gpu=10, drop_last=True), # Batch size = 10, workers_per_gpu = 4
val_dataloader=dict(samples_per_gpu=6, workers_per_gpu=1), # Batch size = 6, workers_per_gpu = 1
test_dataloader=dict(workers_per_gpu=16), # Batch size = 8, workers_per_gpu = 16
...
)
# Evaluation
evaluation = dict(interval=1, by_epoch=True) # Evaluate the model every epoch
# Saving and Logging
checkpoint_config = dict(interval=1) # Save a checkpoint every epoch
log_config = dict(
interval=5, # Print out the model's performance every 5 iterations
hooks=[
dict(type='TextLoggerHook')
])
# Optimizer
optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001) # Supports all optimizers in PyTorch and shares the same parameters
optimizer_config = dict(grad_clip=None) # Parameters for the optimizer hook. See https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/optimizer.py for implementation details
# Learning policy
lr_config = dict(policy='poly', power=0.9, min_lr=1e-7, by_epoch=True)
Testing¶
We introduce the way to test pretrained models on datasets here.
Testing on a Single GPU¶
You can use tools/test.py
to perform single CPU/GPU inference. For example, to evaluate DBNet on IC15: (You can download pretrained models from Model Zoo):
./tools/dist_test.sh configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py dbnet_r18_fpnc_sbn_1200e_icdar2015_20210329-ba3ab597.pth --eval hmean-iou
And here is the full usage of the script:
python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [ARGS]
Note
By default, MMOCR prefers GPU(s) to CPU. If you want to test a model on CPU, please empty CUDA_VISIBLE_DEVICES
or set it to -1 to make GPU(s) invisible to the program. Note that running CPU tests requires MMCV >= 1.4.4.
CUDA_VISIBLE_DEVICES= python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [ARGS]
ARGS | Type | Description |
---|---|---|
--out |
str | Output result file in pickle format. |
--fuse-conv-bn |
bool | Path to the custom config of the selected det model. |
--format-only |
bool | Format the output results without performing evaluation. It is useful when you want to format the results to a specific format and submit them to the test server. |
--gpu-id |
int | GPU id to use. Only applicable to non-distributed training. |
--eval |
'hmean-ic13', 'hmean-iou', 'acc', 'macro-f1' | The evaluation metrics. Options: 'hmean-ic13', 'hmean-iou' for text detection tasks, 'acc' for text recognition tasks, and 'macro-f1' for key information extraction tasks. |
--show |
bool | Whether to show results. |
--show-dir |
str | Directory where the output images will be saved. |
--show-score-thr |
float | Score threshold (default: 0.3). |
--gpu-collect |
bool | Whether to use gpu to collect results. |
--tmpdir |
str | The tmp directory used for collecting results from multiple workers, available when gpu-collect is not specified. |
--cfg-options |
str | Override some settings in the used config, the key-value pair in xxx=yyy format will be merged into the config file. If the value to be overwritten is a list, it should be of the form of either key="[a,b]" or key=a,b. The argument also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]". Note that the quotation marks are necessary and that no white space is allowed. |
--eval-options |
str | Custom options for evaluation, the key-value pair in xxx=yyy format will be kwargs for dataset.evaluate() function. |
--launcher |
'none', 'pytorch', 'slurm', 'mpi' | Options for job launcher. |
Testing on Multiple GPUs¶
MMOCR implements distributed testing with MMDistributedDataParallel
.
You can use the following command to test a dataset with multiple GPUs.
[PORT={PORT}] ./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} [PY_ARGS]
Arguments | Type | Description |
---|---|---|
PORT |
int | The master port that will be used by the machine with rank 0. Defaults to 29500. |
CONFIG_FILE |
str | The path to config. |
CHECKPOINT_FILE |
str | The path to the checkpoint. |
GPU_NUM |
int | The number of GPUs to be used per node. Defaults to 8. |
PY_ARGS |
str | Arguments to be parsed by tools/test.py . |
For example,
./tools/dist_test.sh configs/example_config.py work_dirs/example_exp/example_model_20200202.pth 1 --eval hmean-iou
Testing on Multiple Machines¶
You can launch a task on multiple machines connected to the same network.
NNODES=${NNODES} NODE_RANK=${NODE_RANK} PORT=${MASTER_PORT} MASTER_ADDR=${MASTER_ADDR} ./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} [PY_ARGS]
Arguments | Type | Description |
---|---|---|
NNODES |
int | The number of nodes. |
NODE_RANK |
int | The rank of current node. |
PORT |
int | The master port that will be used by rank 0 node. Defaults to 29500. |
MASTER_ADDR |
str | The address of rank 0 node. Defaults to "127.0.0.1". |
CONFIG_FILE |
str | The path to config. |
CHECKPOINT_FILE |
str | The path to the checkpoint. |
GPU_NUM |
int | The number of GPUs to be used per node. Defaults to 8. |
PY_ARGS |
str | Arguments to be parsed by tools/test.py . |
Note
MMOCR relies on torch.distributed package for distributed testing. Find more information at PyTorch’s launch utility.
Say that you want to launch a job on two machines. On the first machine:
NNODES=2 NODE_RANK=0 PORT=${MASTER_PORT} MASTER_ADDR=${MASTER_ADDR} ./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} [PY_ARGS]
On the second machine:
NNODES=2 NODE_RANK=1 PORT=${MASTER_PORT} MASTER_ADDR=${MASTER_ADDR} ./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} [PY_ARGS]
Note
The speed of the network could be the bottleneck of testing.
Testing with Slurm¶
If you run MMOCR on a cluster managed with Slurm, you can use the script tools/slurm_test.sh
.
[GPUS=${GPUS}] [GPUS_PER_NODE=${GPUS_PER_NODE}] [SRUN_ARGS=${SRUN_ARGS}] ./tools/slurm_test.sh ${PARTITION} ${JOB_NAME} ${CONFIG_FILE} ${CHECKPOINT_FILE} [PY_ARGS]
Arguments | Type | Description |
---|---|---|
GPUS |
int | The number of GPUs to be used by this task. Defaults to 8. |
GPUS_PER_NODE |
int | The number of GPUs to be allocated per node. Defaults to 8. |
SRUN_ARGS |
str | Arguments to be parsed by srun. Available options can be found here. |
PY_ARGS |
str | Arguments to be parsed by tools/test.py . |
Here is an example of using 8 GPUs to test an example model on the ‘dev’ partition with job name ‘test_job’.
GPUS=8 ./tools/slurm_test.sh dev test_job configs/example_config.py work_dirs/example_exp/example_model_20200202.pth --eval hmean-iou
Batch Testing¶
By default, MMOCR tests the model image by image. For faster inference, you may change data.val_dataloader.samples_per_gpu
and data.test_dataloader.samples_per_gpu
in the config. For example,
data = dict(
...
val_dataloader=dict(samples_per_gpu=16),
test_dataloader=dict(samples_per_gpu=16),
...
)
will test the model with 16 images in a batch.
Warning
Batch testing may incur performance decrease of the model due to the different behavior of the data preprocessing pipeline.
Deployment¶
We provide deployment tools under tools/deployment
directory.
Convert to ONNX (experimental)¶
We provide a script to convert the model to ONNX format. The converted model could be visualized by tools like Netron. Besides, we also support comparing the output results between PyTorch and ONNX model.
python tools/deployment/pytorch2onnx.py
${MODEL_CONFIG_PATH} \
${MODEL_CKPT_PATH} \
${MODEL_TYPE} \
${IMAGE_PATH} \
--output-file ${OUTPUT_FILE} \
--device-id ${DEVICE_ID} \
--opset-version ${OPSET_VERSION} \
--verify \
--verbose \
--show \
--dynamic-export
Description of arguments:
ARGS | Type | Description |
---|---|---|
model_config |
str | The path to a model config file. |
model_ckpt |
str | The path to a model checkpoint file. |
model_type |
'recog', 'det' | The model type of the config file. |
image_path |
str | The path to input image file. |
--output-file |
str | The path to output ONNX model. Defaults to tmp.onnx . |
--device-id |
int | Which GPU to use. Defaults to 0. |
--opset-version |
int | ONNX opset version. Defaults to 11. |
--verify |
bool | Determines whether to verify the correctness of an exported model. Defaults to False . |
--verbose |
bool | Determines whether to print the architecture of the exported model. Defaults to False . |
--show |
bool | Determines whether to visualize outputs of ONNXRuntime and PyTorch. Defaults to False . |
--dynamic-export |
bool | Determines whether to export ONNX model with dynamic input and output shapes. Defaults to False . |
Note
This tool is still experimental. For now, some customized operators are not supported, and we only support a subset of detection and recognition algorithms.
List of supported models exportable to ONNX¶
The table below lists the models that are guaranteed to be exportable to ONNX and runnable in ONNX Runtime.
Model | Config | Dynamic Shape | Batch Inference | Note |
---|---|---|---|---|
DBNet | dbnet_r18_fpnc_1200e_icdar2015.py | Y | N | |
PSENet | psenet_r50_fpnf_600e_ctw1500.py | Y | Y | |
PSENet | psenet_r50_fpnf_600e_icdar2015.py | Y | Y | |
PANet | panet_r18_fpem_ffm_600e_ctw1500.py | Y | Y | |
PANet | panet_r18_fpem_ffm_600e_icdar2015.py | Y | Y | |
CRNN | crnn_academic_dataset.py | Y | Y | CRNN only accepts input with height 32 |
Note
All models above are tested with PyTorch==1.8.1 and onnxruntime-gpu == 1.8.1
If you meet any problem with the listed models above, please create an issue and it would be taken care of soon.
Because this feature is experimental and may change fast, please always try with the latest
mmcv
andmmocr
.
Convert ONNX to TensorRT (experimental)¶
We also provide a script to convert ONNX model to TensorRT format. Besides, we support comparing the output results between ONNX and TensorRT model.
python tools/deployment/onnx2tensorrt.py
${MODEL_CONFIG_PATH} \
${MODEL_TYPE} \
${IMAGE_PATH} \
${ONNX_FILE} \
--trt-file ${OUT_TENSORRT} \
--max-shape INT INT INT INT \
--min-shape INT INT INT INT \
--workspace-size INT \
--fp16 \
--verify \
--show \
--verbose
Description of arguments:
ARGS | Type | Description |
---|---|---|
model_config |
str | The path to a model config file. |
model_type |
'recog', 'det' | The model type of the config file. |
image_path |
str | The path to input image file. |
onnx_file |
str | The path to input ONNX file. |
--trt-file |
str | The path of output TensorRT model. Defaults to tmp.trt . |
--max-shape |
int * 4 | Maximum shape of model input. |
--min-shape |
int * 4 | Minimum shape of model input. |
--workspace-size |
int | Max workspace size in GiB. Defaults to 1. |
--fp16 |
bool | Determines whether to export TensorRT with fp16 mode. Defaults to False . |
--verify |
bool | Determines whether to verify the correctness of an exported model. Defaults to False . |
--show |
bool | Determines whether to show the output of ONNX and TensorRT. Defaults to False . |
--verbose |
bool | Determines whether to verbose logging messages while creating TensorRT engine. Defaults to False . |
Note
This tool is still experimental. For now, some customized operators are not supported, and we only support a subset of detection and recognition algorithms.
List of supported models exportable to TensorRT¶
The table below lists the models that are guaranteed to be exportable to TensorRT engine and runnable in TensorRT.
Model | Config | Dynamic Shape | Batch Inference | Note |
---|---|---|---|---|
DBNet | dbnet_r18_fpnc_1200e_icdar2015.py | Y | N | |
PSENet | psenet_r50_fpnf_600e_ctw1500.py | Y | Y | |
PSENet | psenet_r50_fpnf_600e_icdar2015.py | Y | Y | |
PANet | panet_r18_fpem_ffm_600e_ctw1500.py | Y | Y | |
PANet | panet_r18_fpem_ffm_600e_icdar2015.py | Y | Y | |
CRNN | crnn_academic_dataset.py | Y | Y | CRNN only accepts input with height 32 |
Note
All models above are tested with PyTorch==1.8.1, onnxruntime-gpu==1.8.1 and tensorrt==7.2.1.6
If you meet any problem with the listed models above, please create an issue and it would be taken care of soon.
Because this feature is experimental and may change fast, please always try with the latest
mmcv
andmmocr
.
Evaluate ONNX and TensorRT Models (experimental)¶
We provide methods to evaluate TensorRT and ONNX models in tools/deployment/deploy_test.py
.
Prerequisite¶
To evaluate ONNX and TensorRT models, ONNX, ONNXRuntime and TensorRT should be installed first. Install mmcv-full
with ONNXRuntime custom ops and TensorRT plugins follow ONNXRuntime in mmcv and TensorRT plugin in mmcv.
Usage¶
python tools/deploy_test.py \
${CONFIG_FILE} \
${MODEL_PATH} \
${MODEL_TYPE} \
${BACKEND} \
--eval ${METRICS} \
--device ${DEVICE}
Description of all arguments¶
ARGS | Type | Description |
---|---|---|
model_config |
str | The path to a model config file. |
model_file |
str | The path to a TensorRT or an ONNX model file. |
model_type |
'recog', 'det' | Detection or recognition model to deploy. |
backend |
'TensorRT', 'ONNXRuntime' | The backend for testing. |
--eval |
'acc', 'hmean-iou' | The evaluation metrics. 'acc' for recognition models, 'hmean-iou' for detection models. |
--device |
str | Device for evaluation. Defaults to cuda:0 . |
Results and Models¶
Model | Config | Dataset | Metric | PyTorch | ONNX Runtime | TensorRT FP32 | TensorRT FP16 |
---|---|---|---|---|---|---|---|
DBNet | dbnet_r18_fpnc_1200e_icdar2015.py |
icdar2015 | Recall |
0.731 | 0.731 | 0.678 | 0.679 |
Precision | 0.871 | 0.871 | 0.844 | 0.842 | |||
Hmean | 0.795 | 0.795 | 0.752 | 0.752 | |||
DBNet* | dbnet_r18_fpnc_1200e_icdar2015.py |
icdar2015 | Recall |
0.720 | 0.720 | 0.720 | 0.718 |
Precision | 0.868 | 0.868 | 0.868 | 0.868 | |||
Hmean | 0.787 | 0.787 | 0.787 | 0.786 | |||
PSENet | psenet_r50_fpnf_600e_icdar2015.py |
icdar2015 | Recall |
0.753 | 0.753 | 0.753 | 0.752 |
Precision | 0.867 | 0.867 | 0.867 | 0.867 | |||
Hmean | 0.806 | 0.806 | 0.806 | 0.805 | |||
PANet | panet_r18_fpem_ffm_600e_icdar2015.py |
icdar2015 | Recall |
0.740 | 0.740 | 0.687 | N/A |
Precision | 0.860 | 0.860 | 0.815 | N/A | |||
Hmean | 0.796 | 0.796 | 0.746 | N/A | |||
PANet* | panet_r18_fpem_ffm_600e_icdar2015.py |
icdar2015 | Recall |
0.736 | 0.736 | 0.736 | N/A |
Precision | 0.857 | 0.857 | 0.857 | N/A | |||
Hmean | 0.792 | 0.792 | 0.792 | N/A | |||
CRNN | crnn_academic_dataset.py |
IIIT5K | Acc | 0.806 | 0.806 | 0.806 | 0.806 |
Note
TensorRT upsampling operation is a little different from PyTorch. For DBNet and PANet, we suggest replacing upsampling operations with the nearest mode to operations with bilinear mode. Here for PANet, here and here for DBNet. As is shown in the above table, networks with tag * mean the upsampling mode is changed.
Note that changing upsampling mode reduces less performance compared with using the nearest mode. However, the weights of networks are trained through the nearest mode. To pursue the best performance, using bilinear mode for both training and TensorRT deployment is recommended.
All ONNX and TensorRT models are evaluated with dynamic shapes on the datasets, and images are preprocessed according to the original config file.
This tool is still experimental, and we only support a subset of detection and recognition algorithms for now.
C++ Inference example with OpenCV¶
The example below is tested with Visual Studio 2019 as console application, CPU inference only.
Prerequisites¶
Project should use OpenCV (tested with version 4.5.4), ONNX Runtime NuGet package (version 1.9.0).
Download DBNet_r18 detector and SATRN_small recognizer models from our Model Zoo, and export them with the following python commands (you may change the paths accordingly):
python3.9 ../mmocr/tools/deployment/pytorch2onnx.py --verify --output-file detector.onnx ../mmocr/configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py ./dbnet_r18_fpnc_sbn_1200e_icdar2015_20210329-ba3ab597.pth --dynamic-export det ./sample_big_image_eg_1920x1080.png
python3.9 ../mmocr/tools/deployment/pytorch2onnx.py --opset 14 --verify --output-file recognizer.onnx ../mmocr/configs/textrecog/satrn/satrn_small.py ./satrn_small_20211009-2cf13355.pth recog ./sample_small_image_eg_200x50.png
Note
Be aware, while exported
detector.onnx
file is relatively small (about 50 Mb),recognizer.onnx
is pretty big (more than 600 Mb).DBNet_r18 can use ONNX opset 11, SATRN_small can be exported with opset 14.
Warning
Be sure, that verifications of both models are successful - look through the export messages.
Example¶
Example usage of exported models with C++ is in the code below (don’t forget to change paths to *.onnx files). It’s applicable to these two models only, other models have another preprocessing and postprocessing logics.
#include <iostream>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/dnn.hpp>
#include <onnxruntime_cxx_api.h>
#pragma comment(lib, "onnxruntime.lib")
// DB_r18
class Detector {
public:
Detector(const std::string& model_path) {
session = Ort::Session{ env, std::wstring(model_path.begin(), model_path.end()).c_str(), Ort::SessionOptions{nullptr} };
}
std::vector<cv::Rect> inference(const cv::Mat& original, float threshold = 0.3f) {
cv::Size original_size = original.size();
const char* input_names[] = { "input" };
const char* output_names[] = { "output" };
std::array<int64_t, 4> input_shape{ 1, 3, height, width };
cv::Mat image = cv::Mat::zeros(cv::Size(width, height), original.type());
cv::resize(original, image, cv::Size(width, height), 0, 0, cv::INTER_AREA);
image.convertTo(image, CV_32FC3);
cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
image = (image - cv::Scalar(123.675f, 116.28f, 103.53f)) / cv::Scalar(58.395f, 57.12f, 57.375f);
cv::Mat blob = cv::dnn::blobFromImage(image);
auto memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
Ort::Value input_tensor = Ort::Value::CreateTensor<float>(memory_info, (float*)blob.data, blob.total(), input_shape.data(), input_shape.size());
std::vector<Ort::Value> output_tensor = session.Run(Ort::RunOptions{ nullptr }, input_names, &input_tensor, 1, output_names, 1);
int sizes[] = { 1, 3, height, width };
cv::Mat output(4, sizes, CV_32F, output_tensor.front().GetTensorMutableData<float>());
std::vector<cv::Mat> images;
cv::dnn::imagesFromBlob(output, images);
std::vector<cv::Rect> areas = get_detected(images[0], threshold);
std::vector<cv::Rect> results;
float x_ratio = original_size.width / (float)width;
float y_ratio = original_size.height / (float)height;
for (int index = 0; index < areas.size(); ++index) {
cv::Rect box = areas[index];
box.x = int(box.x * x_ratio);
box.width = int(box.width * x_ratio);
box.y = int(box.y * y_ratio);
box.height = int(box.height * y_ratio);
results.push_back(box);
}
return results;
}
private:
Ort::Env env;
Ort::Session session{ nullptr };
const int width = 1312, height = 736;
cv::Rect expand_box(const cv::Rect& original, int addition = 5) {
cv::Rect box(original);
box.x = std::max(0, box.x - addition);
box.y = std::max(0, box.y - addition);
box.width = (box.x + box.width + addition * 2 > width) ? (width - box.x) : (box.width + addition * 2);
box.height = (box.y + box.height + addition * 2) > height ? (height - box.y) : (box.height + addition * 2);
return box;
}
std::vector<cv::Rect> get_detected(const cv::Mat& output, float threshold) {
cv::Mat text_mask = cv::Mat::zeros(height, width, CV_32F);
std::vector<cv::Mat> maps;
cv::split(output, maps);
cv::Mat proba_map = maps[0];
cv::threshold(proba_map, text_mask, threshold, 1.0f, cv::THRESH_BINARY);
cv::multiply(text_mask, 255, text_mask);
text_mask.convertTo(text_mask, CV_8U);
std::vector<std::vector<cv::Point>> contours;
cv::findContours(text_mask, contours, cv::RETR_EXTERNAL, cv::CHAIN_APPROX_SIMPLE);
std::vector<cv::Rect> boxes;
for (int index = 0; index < contours.size(); ++index) {
cv::Rect box = expand_box(cv::boundingRect(contours[index]));
boxes.push_back(box);
}
return boxes;
}
};
// SATRN_small
class Recognizer {
public:
Recognizer(const std::string& model_path) {
session = Ort::Session{ env, std::wstring(model_path.begin(), model_path.end()).c_str(), Ort::SessionOptions{nullptr} };
}
std::string inference(const cv::Mat& original) {
const char* input_names[] = { "input" };
const char* output_names[] = { "output" };
std::array<int64_t, 4> input_shape{ 1, 3, height, width };
cv::Mat image;
cv::resize(original, image, cv::Size(width, height), 0, 0, cv::INTER_AREA);
image.convertTo(image, CV_32FC3);
cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
image = (image / 255.0f - cv::Scalar(0.485f, 0.456f, 0.406f)) / cv::Scalar(0.229f, 0.224f, 0.225f);
cv::Mat blob = cv::dnn::blobFromImage(image);
auto memory_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
Ort::Value input_tensor = Ort::Value::CreateTensor<float>(memory_info, (float*)blob.data, blob.total(), input_shape.data(), input_shape.size());
std::vector<Ort::Value> output_tensor = session.Run(Ort::RunOptions{ nullptr }, input_names, &input_tensor, 1, output_names, 1);
int sequence_length = 25;
std::string dictionary = "0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!\"#$%&'()*+,-./:;<=>?@[\\]_`~";
int characters = dictionary.length() + 2; // EOS + UNK
std::vector<int> max_indices;
for (int outer = 0; outer < sequence_length; ++outer) {
int character_index = -1;
float character_value = 0;
for (int inner = 0; inner < characters; ++inner) {
int counter = outer * characters + inner;
float value = output_tensor[0].GetTensorMutableData<float>()[counter];
if (value > character_value) {
character_value = value;
character_index = inner;
}
}
max_indices.push_back(character_index);
}
std::string recognized;
for (int index = 0; index < max_indices.size(); ++index) {
if (max_indices[index] == dictionary.length()) {
continue; // unk
}
if (max_indices[index] == dictionary.length() + 1) {
break; // eos
}
recognized += dictionary[max_indices[index]];
}
return recognized;
}
private:
Ort::Env env;
Ort::Session session{ nullptr };
const int height = 32;
const int width = 100;
};
int main(int argc, const char* argv[]) {
if (argc < 2) {
std::cout << "Usage: this_executable.exe c:/path/to/image.png" << std::endl;
return 0;
}
std::chrono::steady_clock::time_point begin = std::chrono::steady_clock::now();
std::cout << "Loading models..." << std::endl;
Detector detector("d:/path/to/detector.onnx");
Recognizer recognizer("d:/path/to/recognizer.onnx");
std::chrono::steady_clock::time_point end = std::chrono::steady_clock::now();
std::cout << "Loading models done in " << std::chrono::duration_cast<std::chrono::milliseconds>(end - begin).count() << " ms" << std::endl;
cv::Mat image = cv::imread(argv[1], cv::IMREAD_COLOR);
begin = std::chrono::steady_clock::now();
std::vector<cv::Rect> detections = detector.inference(image);
for (int index = 0; index < detections.size(); ++index) {
cv::Mat roi = image(detections[index]);
std::string text = recognizer.inference(roi);
cv::rectangle(image, detections[index], cv::Scalar(255, 255, 255), 2);
cv::putText(image, text, cv::Point(detections[index].x, detections[index].y - 10), cv::FONT_HERSHEY_COMPLEX, 0.4, cv::Scalar(255, 255, 255));
}
end = std::chrono::steady_clock::now();
std::cout << "Inference time (with drawing): " << std::chrono::duration_cast<std::chrono::milliseconds>(end - begin).count() << " ms" << std::endl;
cv::imshow("Results", image);
cv::waitKey(0);
return 0;
}
The output should look something like this.
Loading models...
Loading models done in 5715 ms
Inference time (with drawing): 3349 ms
And the sample result should look something like this.
Model Serving¶
MMOCR
provides some utilities that facilitate the model serving process.
Here is a quick walkthrough of necessary steps that let the models to serve through an API.
Install TorchServe¶
You can follow the steps on the official website to install TorchServe
and
torch-model-archiver
.
Convert model from MMOCR to TorchServe¶
We provide a handy tool to convert any .pth
model into .mar
model
for TorchServe.
python tools/deployment/mmocr2torchserve.py ${CONFIG_FILE} ${CHECKPOINT_FILE} \
--output-folder ${MODEL_STORE} \
--model-name ${MODEL_NAME}
Note
${MODEL_STORE} needs to be an absolute path to a folder.
For example:
python tools/deployment/mmocr2torchserve.py \
configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py \
checkpoints/dbnet_r18_fpnc_1200e_icdar2015.pth \
--output-folder ./checkpoints \
--model-name dbnet
Start Serving¶
From your Local Machine¶
Getting your models prepared, the next step is to start the service with a one-line command:
# To load all the models in ./checkpoints
torchserve --start --model-store ./checkpoints --models all
# Or, if you only want one model to serve, say dbnet
torchserve --start --model-store ./checkpoints --models dbnet=dbnet.mar
Then you can access inference, management and metrics services through TorchServe’s REST API. You can find their usages in TorchServe REST API.
Service | Address |
---|---|
Inference | http://127.0.0.1:8080 |
Management | http://127.0.0.1:8081 |
Metrics | http://127.0.0.1:8082 |
Note
By default, TorchServe binds port number 8080
, 8081
and 8082
to its services.
You can change such behavior by modifying and saving the contents below to config.properties
, and running TorchServe with option --ts-config config.properties
.
inference_address=http://0.0.0.0:8080
management_address=http://0.0.0.0:8081
metrics_address=http://0.0.0.0:8082
number_of_netty_threads=32
job_queue_size=1000
model_store=/home/model-server/model-store
From Docker¶
A better alternative to serve your models is through Docker. We provide a Dockerfile that frees you from those tedious and error-prone environmental setup steps.
Build mmocr-serve
Docker image¶
docker build -t mmocr-serve:latest docker/serve/
Run mmocr-serve
with Docker¶
In order to run Docker in GPU, you need to install nvidia-docker; or you can omit the --gpus
argument for a CPU-only session.
The command below will run mmocr-serve
with a gpu, bind the ports of 8080
(inference),
8081
(management) and 8082
(metrics) from container to 127.0.0.1
, and mount
the checkpoint folder ./checkpoints
from the host machine to /home/model-server/model-store
of the container. For more information, please check the official docs for running TorchServe with docker.
docker run --rm \
--cpus 8 \
--gpus device=0 \
-p8080:8080 -p8081:8081 -p8082:8082 \
--mount type=bind,source=`realpath ./checkpoints`,target=/home/model-server/model-store \
mmocr-serve:latest
Note
realpath ./checkpoints
points to the absolute path of “./checkpoints”, and you can replace it with the absolute path where you store torchserve models.
Upon running the docker, you can access inference, management and metrics services through TorchServe’s REST API. You can find their usages in TorchServe REST API.
Service | Address |
---|---|
Inference | http://127.0.0.1:8080 |
Management | http://127.0.0.1:8081 |
Metrics | http://127.0.0.1:8082 |
4. Test deployment¶
Inference API allows user to post an image to a model and returns the prediction result.
curl http://127.0.0.1:8080/predictions/${MODEL_NAME} -T demo/demo_text_det.jpg
For example,
curl http://127.0.0.1:8080/predictions/dbnet -T demo/demo_text_det.jpg
For detection models, you should obtain a json with an object named boundary_result
. Each array inside has float numbers representing x, y
coordinates of boundary vertices in clockwise order, and the last float number as the
confidence score.
{
"boundary_result": [
[
221.18990004062653,
226.875,
221.18990004062653,
212.625,
244.05868631601334,
212.625,
244.05868631601334,
226.875,
0.80883354575186
]
]
}
For recognition models, the response should look like:
{
"text": "sier",
"score": 0.5247521847486496
}
And you can use test_torchserve.py
to compare result of TorchServe and PyTorch by visualizing them.
python tools/deployment/test_torchserve.py ${IMAGE_FILE} ${CONFIG_FILE} ${CHECKPOINT_FILE} ${MODEL_NAME}
[--inference-addr ${INFERENCE_ADDR}] [--device ${DEVICE}]
Example:
python tools/deployment/test_torchserve.py \
demo/demo_text_det.jpg \
configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py \
checkpoints/dbnet_r18_fpnc_1200e_icdar2015.pth \
dbnet
Learn about Configs¶
We incorporate modular and inheritance design into our config system, which is convenient to conduct various experiments.
If you wish to inspect the config file, you may run python tools/misc/print_config.py /PATH/TO/CONFIG
to see the complete config.
Modify config through script arguments¶
When submitting jobs using “tools/train.py” or “tools/test.py”, you may specify --cfg-options
to in-place modify the config.
Update config keys of dict chains.
The config options can be specified following the order of the dict keys in the original config. For example,
--cfg-options model.backbone.norm_eval=False
changes the all BN modules in model backbones totrain
mode.Update keys inside a list of configs.
Some config dicts are composed as a list in your config. For example, the training pipeline
data.train.pipeline
is normally a list e.g.[dict(type='LoadImageFromFile'), ...]
. If you want to change'LoadImageFromFile'
to'LoadImageFromNdarry'
in the pipeline, you may specify--cfg-options data.train.pipeline.0.type=LoadImageFromNdarry
.Update values of list/tuples.
If the value to be updated is a list or a tuple. For example, the config file normally sets
workflow=[('train', 1)]
. If you want to change this key, you may specify--cfg-options workflow="[(train,1),(val,1)]"
. Note that the quotation mark ” is necessary to support list/tuple data types, and that NO white space is allowed inside the quotation marks in the specified value.
Config Name Style¶
We follow the below style to name full config files (configs/TASK/*.py
). Contributors are advised to follow the same style.
{model}_[ARCHITECTURE]_[schedule]_{dataset}.py
{xxx}
is required field and [yyy]
is optional.
{model}
: model type likedbnet
,crnn
, etc.[ARCHITECTURE]
: expands some invoked modules following the order of data flow, and the content depends on the model framework. The following examples show how it is generally expanded.For text detection tasks, key information tasks, and SegOCR in text recognition task:
{model}_[backbone]_[neck]_[schedule]_{dataset}.py
For other text recognition tasks,
{model}_[backbone]_[encoder]_[decoder]_[schedule]_{dataset}.py
Note thatbackbone
,neck
,encoder
,decoder
are the names of modules, e.g.r50
,fpnocr
, etc.
{schedule}
: training schedule. For instance,1200e
denotes 1200 epochs.{dataset}
: dataset. It can either be the name of a dataset (icdar2015
), or a collection of datasets for brevity (e.g.academic
usually refers to a common practice in academia, which uses MJSynth + SynthText as training set, and IIIT5K, SVT, IC13, IC15, SVTP and CT80 as test set).
Most configs are composed of basic primitive configs in configs/_base_
, where each primitive config in different subdirectory has a slightly different name style. We present them as follows.
det_datasets, recog_datasets:
{dataset_name(s)}_[train|test].py
. If [train|test] is not specified, the config should contain both training and test set.There are two exceptions: toy_data.py and seg_toy_data.py. In recog_datasets, the first one works for most while the second one contains character level annotations and works for seg baseline only as of Dec 2021.
det_models, recog_models:
{model}_[ARCHITECTURE].py
.det_pipelines, recog_pipelines:
{model}_pipeline.py
.schedules:
schedule_{optimizer}_{num_epochs}e.py
.
Config Structure¶
For better config reusability, we break many of reusable sections of configs into configs/_base_
. Now the directory tree of configs/_base_
is organized as follows:
_base_
├── det_datasets
├── det_models
├── det_pipelines
├── recog_datasets
├── recog_models
├── recog_pipelines
└── schedules
These primitive configs are categorized by their roles in a complete config. Most of model configs are making full use of primitive configs by including them as parts of _base_
section. For example, dbnet_r18_fpnc_1200e_icdar2015.py takes five primitive configs from _base_
:
_base_ = [
'../../_base_/default_runtime.py',
'../../_base_/schedules/schedule_sgd_1200e.py',
'../../_base_/det_models/dbnet_r18_fpnc.py',
'../../_base_/det_datasets/icdar2015.py',
'../../_base_/det_pipelines/dbnet_pipeline.py'
]
From these configs’ names we can roughly know this config trains dbnet_r18_fpnc with sgd optimizer in 1200 epochs. It uses the origin dbnet pipeline and icdar2015 as the dataset. We encourage users to follow and take advantage of this convention to organize the config clearly and facilitate fair comparison across different primitive configurations as well as models.
Please refer to mmcv for detailed documentation.
Config File Structure¶
Model¶
The parameter "model"
is a python dictionary in the configuration file, which mainly includes information such as network structure and loss function.
Note
The ‘type’ in the configuration file is not a constructed parameter, but a class name.
Note
We can also use models from MMDetection by adding mmdet.
prefix to type name, or from other OpenMMLab projects in a similar way if their backbones are registered in registries.
Text Detection / Text Recognition / Key Information Extraction Model¶
backbone
: Backbone configs. Common Backbones, TextRecog Backbonesneck
: Neck network name. TextDet Necks, TextRecog Necks.bbox_head
: Head network name. Applicable to text detection, key information models and some text recognition models. TextDet Heads, TextRecog Heads, KIE Heads.loss
: Loss function type. TextDet Losses, KIE Lossespostprocessor
: (TextDet only) Postprocess type. TextDet Postprocessors
Text Recognition / Named Entity Extraction Model¶
encoder
: Encoder configs. TextRecog Encodersdecoder
: Decoder configs. Applicable to text recognition models. TextRecog Decodersloss
: Loss configs. Applicable to some text recognition models. TextRecog Losseslabel_convertor
: Convert outputs between text, index and tensor. Applicable to text recognition models. Label Convertorsmax_seq_len
: The maximum sequence length of recognition results. Applicable to text recognition models.
Data & Pipeline¶
The parameter "data"
is a python dictionary in the configuration file, which mainly includes information to construct dataloader:
samples_per_gpu
: the BatchSize of each GPU when building the dataloaderworkers_per_gpu
: the number of threads per GPU when building dataloadertrain | val | test
: config to construct datasettype
: Dataset name. Check dataset types for supported datasets.
The parameter evaluation
is also a dictionary, which is the configuration information of evaluation hook
, mainly including evaluation interval, evaluation index, etc.
# dataset settings
dataset_type = 'IcdarDataset' # dataset name,
data_root = 'data/icdar2015' # dataset root
img_norm_cfg = dict( # Image normalization config to normalize the input images
mean=[123.675, 116.28, 103.53], # Mean values used to pre-training the pre-trained backbone models
std=[58.395, 57.12, 57.375], # Standard variance used to pre-training the pre-trained backbone models
to_rgb=True) # Whether to invert the color channel, rgb2bgr or bgr2rgb.
# train data pipeline
train_pipeline = [ # Training pipeline
dict(type='LoadImageFromFile'), # First pipeline to load images from file path
dict(
type='LoadAnnotations', # Second pipeline to load annotations for current image
with_bbox=True, # Whether to use bounding box, True for detection
with_mask=True, # Whether to use instance mask, True for instance segmentation
poly2mask=False), # Whether to convert the polygon mask to instance mask, set False for acceleration and to save memory
dict(
type='Resize', # Augmentation pipeline that resize the images and their annotations
img_scale=(1333, 800), # The largest scale of image
keep_ratio=True
), # whether to keep the ratio between height and width.
dict(
type='RandomFlip', # Augmentation pipeline that flip the images and their annotations
flip_ratio=0.5), # The ratio or probability to flip
dict(
type='Normalize', # Augmentation pipeline that normalize the input images
mean=[123.675, 116.28, 103.53], # These keys are the same of img_norm_cfg since the
std=[58.395, 57.12, 57.375], # keys of img_norm_cfg are used here as arguments
to_rgb=True),
dict(
type='Pad', # Padding config
size_divisor=32), # The number the padded images should be divisible
dict(type='DefaultFormatBundle'), # Default format bundle to gather data in the pipeline
dict(
type='Collect', # Pipeline that decides which keys in the data should be passed to the detector
keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks'])
]
test_pipeline = [
dict(type='LoadImageFromFile'), # First pipeline to load images from file path
dict(
type='MultiScaleFlipAug', # An encapsulation that encapsulates the testing augmentations
img_scale=(1333, 800), # Decides the largest scale for testing, used for the Resize pipeline
flip=False, # Whether to flip images during testing
transforms=[
dict(type='Resize', # Use resize augmentation
keep_ratio=True), # Whether to keep the ratio between height and width, the img_scale set here will be suppressed by the img_scale set above.
dict(type='RandomFlip'), # Thought RandomFlip is added in pipeline, it is not used because flip=False
dict(
type='Normalize', # Normalization config, the values are from img_norm_cfg
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(
type='Pad', # Padding config to pad images divisible by 32.
size_divisor=32),
dict(
type='ImageToTensor', # convert image to tensor
keys=['img']),
dict(
type='Collect', # Collect pipeline that collect necessary keys for testing.
keys=['img'])
])
]
data = dict(
samples_per_gpu=32, # Batch size of a single GPU
workers_per_gpu=2, # Worker to pre-fetch data for each single GPU
train=dict( # train data config
type=dataset_type, # dataset name
ann_file=f'{data_root}/instances_training.json', # Path to annotation file
img_prefix=f'{data_root}/imgs', # Path to images
pipeline=train_pipeline), # train data pipeline
test=dict( # test data config
type=dataset_type,
ann_file=f'{data_root}/instances_test.json', # Path to annotation file
img_prefix=f'{data_root}/imgs', # Path to images
pipeline=test_pipeline))
evaluation = dict( # The config to build the evaluation hook, refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/evaluation/eval_hooks.py#L7 for more details.
interval=1, # Evaluation interval
metric='hmean-iou') # Metrics used during evaluation
Training Schedule¶
Mainly include optimizer settings, optimizer hook
settings, learning rate schedule and runner
settings:
optimizer
: optimizer setting , support all optimizers inpytorch
, refer to related mmcv documentation.optimizer_config
:optimizer hook
configuration file, such as setting gradient limit, refer to related mmcv code.lr_config
: Learning rate scheduler, supports “CosineAnnealing”, “Step”, “Cyclic”, etc. Refer to related mmcv documentation for more options.runner
: Forrunner
, please refer tommcv
forrunner
introduction document.
# The configuration file used to build the optimizer, support all optimizers in PyTorch.
optimizer = dict(type='SGD', # Optimizer type
lr=0.1, # Learning rate of optimizers, see detail usages of the parameters in the documentation of PyTorch
momentum=0.9, # Momentum
weight_decay=0.0001) # Weight decay of SGD
# Config used to build the optimizer hook, refer to https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/optimizer.py#L8 for implementation details.
optimizer_config = dict(grad_clip=None) # Most of the methods do not use gradient clip
# Learning rate scheduler config used to register LrUpdater hook
lr_config = dict(policy='step', # The policy of scheduler, also support CosineAnnealing, Cyclic, etc. Refer to details of supported LrUpdater from https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/lr_updater.py#L9.
step=[30, 60, 90]) # Steps to decay the learning rate
runner = dict(type='EpochBasedRunner', # Type of runner to use (i.e. IterBasedRunner or EpochBasedRunner)
max_epochs=100) # Runner that runs the workflow in total max_epochs. For IterBasedRunner use `max_iters`
Runtime Setting¶
This part mainly includes saving the checkpoint strategy, log configuration, training parameters, breakpoint weight path, working directory, etc..
# Config to set the checkpoint hook, Refer to https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/checkpoint.py for implementation.
checkpoint_config = dict(interval=1) # The save interval is 1
# config to register logger hook
log_config = dict( # Config to register logger hook
interval=50, # Interval to print the log
hooks=[
dict(type='TextLoggerHook', by_epoch=False),
dict(type='TensorboardLoggerHook', by_epoch=False),
dict(type='WandbLoggerHook', by_epoch=False, # The Wandb logger is also supported, It requires `wandb` to be installed.
init_kwargs={
'project': "MMOCR", # Project name in WandB
}), # Check https://docs.wandb.ai/ref/python/init for more init arguments.
# ClearMLLoggerHook, DvcliveLoggerHook, MlflowLoggerHook, NeptuneLoggerHook, PaviLoggerHook, SegmindLoggerHook are also supported based on MMCV implementation.
])
dist_params = dict(backend='nccl') # Parameters to setup distributed training, the port can also be set.
log_level = 'INFO' # The output level of the log.
resume_from = None # Resume checkpoints from a given path, the training will be resumed from the epoch when the checkpoint's is saved.
workflow = [('train', 1)] # Workflow for runner. [('train', 1)] means there is only one workflow and the workflow named 'train' is executed once.
work_dir = 'work_dir' # Directory to save the model checkpoints and logs for the current experiments.
FAQ¶
Ignore some fields in the base configs¶
Sometimes, you may set _delete_=True
to ignore some of fields in base configs.
You may refer to mmcv for simple illustration.
Use intermediate variables in configs¶
Some intermediate variables are used in the configs files, like train_pipeline
/test_pipeline
in datasets.
It’s worth noting that when modifying intermediate variables in the children configs, user need to pass the intermediate variables into corresponding fields again.
For example, we usually want the data path to be a variable so that we
dataset_type = 'IcdarDataset'
data_root = 'data/icdar2015'
train = dict(
type=dataset_type,
ann_file=f'{data_root}/instances_training.json',
img_prefix=f'{data_root}/imgs',
pipeline=None)
test = dict(
type=dataset_type,
ann_file=f'{data_root}/instances_test.json',
img_prefix=f'{data_root}/imgs',
pipeline=None)
Use some fields in the base configs¶
Sometimes, you may refer to some fields in the _base_
config, so as to avoid duplication of definitions. You can refer to mmcv for some more instructions.
This technique has been widely used in MMOCR’s configs, where the main configs refer to the dataset and pipeline defined in base configs by:
train_list = {{_base_.train_list}}
test_list = {{_base_.test_list}}
train_pipeline = {{_base_.train_pipeline}}
test_pipeline = {{_base_.test_pipeline}}
Which assumes that its base configs export datasets and pipelines in a way like:
# base dataset config
dataset_type = 'IcdarDataset'
data_root = 'data/icdar2015'
train = dict(
type=dataset_type,
ann_file=f'{data_root}/instances_training.json',
img_prefix=f'{data_root}/imgs',
pipeline=None)
test = dict(
type=dataset_type,
ann_file=f'{data_root}/instances_test.json',
img_prefix=f'{data_root}/imgs',
pipeline=None)
train_list = [train]
test_list = [test]
# base pipeline config
train_pipeline = dict(...)
test_pipeline = dict(...)
Deprecated train_cfg/test_cfg¶
The train_cfg
and test_cfg
are deprecated in config file, please specify them in the model config. The original config structure is as below.
# deprecated
model = dict(
type=...,
...
)
train_cfg=dict(...)
test_cfg=dict(...)
The migration example is as below.
# recommended
model = dict(
type=...,
...
train_cfg=dict(...),
test_cfg=dict(...),
)
Dataset Types¶
Dataset Wrapper¶
UniformConcatDataset¶
UniformConcatDataset
is a fundamental dataset wrapper in MMOCR which allows users to apply a universal pipeline on multiple datasets without specifying the pipeline for each of them.
Applying a Pipeline on Multiple Datasets¶
For example, to apply train_pipeline
on both train1
and train2
,
data = dict(
...
train=dict(
type='UniformConcatDataset',
datasets=[train1, train2],
pipeline=train_pipeline))
Also, it support applying different pipeline
to different datasets
,
train_list1 = [train1, train2]
train_list2 = [train3, train4]
data = dict(
...
train=dict(
type='UniformConcatDataset',
datasets=[train_list1, train_list2],
pipeline=[train_pipeline1, train_pipeline2]))
Here, train_pipeline1
will be applied to train1
and train2
, and
train_pipeline2
will be applied to train3
and train4
.
Getting Mean Evaluation Scores¶
Evaluating the model on multiple datasets is a common strategy in academia, and the mean score is therefore a critical indicator of the model’s overall performance. By default, UniformConcatDataset
reports mean scores in the form of
mean_{metric_name}
when more than 1 datasets are wrapped. You can customize the behavior by setting
show_mean_scores
in data.val
and data.test
. Choices are 'auto'
(default), True
and False
.
data = dict(
...
val=dict(
type='UniformConcatDataset',
show_mean_scores=True, # always show mean scores
datasets=[train_list],
pipeline=[train_pipeline)
test=dict(
type='UniformConcatDataset',
show_mean_scores=False, # do not show mean scores
datasets=[train_list],
pipeline=[train_pipeline))
Text Detection¶
IcdarDataset¶
Dataset with annotation file in coco-like json format
Example Configuration¶
dataset_type = 'IcdarDataset'
prefix = 'tests/data/toy_dataset/'
test=dict(
type=dataset_type,
ann_file=prefix + 'instances_test.json',
img_prefix=prefix + 'imgs',
pipeline=test_pipeline)
Annotation Format¶
You can check the content of the annotation file in tests/data/toy_dataset/instances_test.json
for an example.
It’s compatible with any annotation file in COCO format defined in MMDetection:
Note
Icdar 2015/2017 and ctw1500 annotations need to be converted into the COCO format following the steps in datasets.md.
Evaluation¶
IcdarDataset
has implemented two evaluation metrics, hmean-iou
and hmean-ic13
, to evaluate the performance of text detection models, where hmean-iou
is the most widely used metric which computes precision, recall and F-score based on IoU between ground truth and prediction.
In particular, filtering predictions with a reasonable score threshold greatly impacts the performance measurement. MMOCR alleviates such hyperparameter effect by sweeping through the hyperparameter space and returns the best performance every evaluation time.
User can tune the searching scheme by passing min_score_thr
, max_score_thr
and step
into the evaluation hook in the config.
For example, with the following configuration, you can evaluate the model’s output on a list of boundary score thresholds [0.1, 0.2, 0.3, 0.4, 0.5] and get the best score from them during training.
evaluation = dict(
interval=100,
metric='hmean-iou',
min_score_thr=0.1,
max_score_thr=0.5,
step=0.1)
During testing, you can change these parameter values by appending them to --eval-options
.
python tools/test.py configs/textdet/dbnet/dbnet_r18_fpnc_1200e_icdar2015.py db_r18.pth --eval hmean-iou --eval-options min_score_thr=0.1 max_score_thr=0.6 step=0.1
Check out our API doc for further explanations on these parameters.
TextDetDataset¶
Dataset with annotation file in line-json txt format
We have designed new types of dataset consisting of loader , backend, and parser to load and parse different types of annotation files.
loader: Load the annotation file. We now have a unified loader,
AnnFileLoader
, which can use differentbackend
to load annotation from txt. The originalHardDiskLoader
andLmdbLoader
will be deprecated.backend: Load annotation from different format and backend.
LmdbAnnFileBackend
: Load annotation from lmdb dataset.HardDiskAnnFileBackend
: Load annotation file with raw hard disks storage backend. The annotation format can be either txt or lmdb.PetrelAnnFileBackend
: Load annotation file with petrel storage backend. The annotation format can be either txt or lmdb.HTTPAnnFileBackend
: Load annotation file with http storage backend. The annotation format can be either txt or lmdb.
parser: Parse the annotation file line-by-line and return with
dict
format. There are two types of parser,LineStrParser
andLineJsonParser
.LineStrParser
: Parse one line in ann file while treating it as a string and separating it to several parts by aseparator
. It can be used on tasks with simple annotation files such as text recognition where each line of the annotation files contains thefilename
andlabel
attribute only.LineJsonParser
: Parse one line in ann file while treating it as a json-string and usingjson.loads
to convert it todict
. It can be used on tasks with complex annotation files such as text detection where each line of the annotation files contains multiple attributes (e.g.filename
,height
,width
,box
,segmentation
,iscrowd
,category_id
, etc.).
Example Configuration¶
dataset_type = 'TextDetDataset'
img_prefix = 'tests/data/toy_dataset/imgs'
test_anno_file = 'tests/data/toy_dataset/instances_test.txt'
test = dict(
type=dataset_type,
img_prefix=img_prefix,
ann_file=test_anno_file,
loader=dict(
type='AnnFileLoader',
repeat=4,
parser=dict(
type='LineJsonParser',
keys=['file_name', 'height', 'width', 'annotations'])),
pipeline=test_pipeline,
test_mode=True)
Annotation Format¶
The results are generated in the same way as the segmentation-based text recognition task above.
You can check the content of the annotation file in tests/data/toy_dataset/instances_test.txt
.
The combination of HardDiskLoader
and LineJsonParser
will return a dict for each file by calling __getitem__
:
{"file_name": "test/img_10.jpg", "height": 720, "width": 1280, "annotations": [{"iscrowd": 1, "category_id": 1, "bbox": [260.0, 138.0, 24.0, 20.0], "segmentation": [[261, 138, 284, 140, 279, 158, 260, 158]]}, {"iscrowd": 0, "category_id": 1, "bbox": [288.0, 138.0, 129.0, 23.0], "segmentation": [[288, 138, 417, 140, 416, 161, 290, 157]]}, {"iscrowd": 0, "category_id": 1, "bbox": [743.0, 145.0, 37.0, 18.0], "segmentation": [[743, 145, 779, 146, 780, 163, 746, 163]]}, {"iscrowd": 0, "category_id": 1, "bbox": [783.0, 129.0, 50.0, 26.0], "segmentation": [[783, 129, 831, 132, 833, 155, 785, 153]]}, {"iscrowd": 1, "category_id": 1, "bbox": [831.0, 133.0, 43.0, 23.0], "segmentation": [[831, 133, 870, 135, 874, 156, 835, 155]]}, {"iscrowd": 1, "category_id": 1, "bbox": [159.0, 204.0, 72.0, 15.0], "segmentation": [[159, 205, 230, 204, 231, 218, 159, 219]]}, {"iscrowd": 1, "category_id": 1, "bbox": [785.0, 158.0, 75.0, 21.0], "segmentation": [[785, 158, 856, 158, 860, 178, 787, 179]]}, {"iscrowd": 1, "category_id": 1, "bbox": [1011.0, 157.0, 68.0, 16.0], "segmentation": [[1011, 157, 1079, 160, 1076, 173, 1011, 170]]}]}
Evaluation¶
TextDetDataset
shares a similar implementation with IcdarDataset
. Please refer to the evaluation section of ‘IcdarDataset’.
Text Recognition¶
OCRDataset¶
Dataset for encoder-decoder based recognizer
It shares a similar architecture with TextDetDataset
. Check out the introduction for details.
Example Configuration¶
dataset_type = 'OCRDataset'
img_prefix = 'tests/data/ocr_toy_dataset/imgs'
train_anno_file = 'tests/data/ocr_toy_dataset/label.txt'
train = dict(
type=dataset_type,
img_prefix=img_prefix,
ann_file=train_anno_file,
loader=dict(
type='AnnFileLoader',
repeat=10,
parser=dict(
type='LineStrParser',
keys=['filename', 'text'],
keys_idx=[0, 1],
separator=' ')),
pipeline=train_pipeline,
test_mode=False)
Optional Arguments:
repeat
: The number of repeated lines in the annotation files. For example, if there are10
lines in the annotation file, settingrepeat=10
will generate a corresponding annotation file with size100
.
Annotation Format¶
You can check the content of the annotation file in tests/data/ocr_toy_dataset/label.txt
.
The combination of HardDiskLoader
and LineStrParser
will return a dict for each file by calling __getitem__
: {'filename': '1223731.jpg', 'text': 'GRAND'}
.
Loading LMDB Datasets¶
We have support for reading annotation files from the full lmdb dataset (with images and annotations). It is now possible to read lmdb datasets commonly used in academia. We have also implemented a new dataset conversion tool, recog2lmdb. It converts the recognition dataset to lmdb format. See PR982 for more details.
Here is an example configuration to load lmdb annotations:
lmdb_root = 'path to lmdb folder'
train = dict(
type='OCRDataset',
img_prefix=lmdb_root,
ann_file=lmdb_root,
loader=dict(
type='AnnFileLoader',
repeat=1,
file_format='lmdb',
parser=dict(
type='LineJsonParser',
keys=['filename', 'text']),
pipeline=None,
test_mode=False)
Evaluation¶
There are six evaluation metrics available for text recognition tasks: word_acc
, word_acc_ignore_case
, word_acc_ignore_case_symbol
, char_recall
, char_precision
and one_minus_ned
. See our API doc for explanations on metrics.
By default, OCRDataset
generates full reports on all the metrics if its evaluation
metric is acc
. Here is an example case for training.
# Configuration
evaluation = dict(interval=1, metric='acc')
# Results
{'0_char_recall': 0.0484, '0_char_precision': 0.6, '0_word_acc': 0.0, '0_word_acc_ignore_case': 0.0, '0_word_acc_ignore_case_symbol': 0.0, '0_1-N.E.D': 0.0525}
Note
‘0_’ prefixes result from UniformConcatDataset
. It’s kept here since MMOCR always wrap UniformConcatDataset
around any datasets.
If you want to conduct the evaluation on a subset of evaluation metrics:
evaluation = dict(interval=1, metric=['word_acc_ignore_case', 'one_minus_ned'])
The result will look like:
{'0_word_acc_ignore_case': 0.0, '0_1-N.E.D': 0.0525}
During testing, you can specify the metrics to evaluate in the command line:
python tools/test.py configs/textrecog/crnn/crnn_toy_dataset.py crnn.pth --eval word_acc_ignore_case one_minus_ned
OCRSegDataset¶
Dataset for segmentation-based recognizer
It shares a similar architecture with TextDetDataset
. Check out the introduction for details.
Example Configuration¶
prefix = 'tests/data/ocr_char_ann_toy_dataset/'
train = dict(
type='OCRSegDataset',
img_prefix=prefix + 'imgs',
ann_file=prefix + 'instances_train.txt',
loader=dict(
type='AnnFileLoader',
repeat=10,
parser=dict(
type='LineJsonParser',
keys=['file_name', 'annotations', 'text'])),
pipeline=train_pipeline,
test_mode=True)
Annotation Format¶
You can check the content of the annotation file in tests/data/ocr_char_ann_toy_dataset/instances_train.txt
.
The combination of HardDiskLoader
and LineJsonParser
will return a dict for each file by calling __getitem__
each time:
{"file_name": "resort_88_101_1.png", "annotations": [{"char_text": "F", "char_box": [11.0, 0.0, 22.0, 0.0, 12.0, 12.0, 0.0, 12.0]}, {"char_text": "r", "char_box": [23.0, 2.0, 31.0, 1.0, 24.0, 11.0, 16.0, 11.0]}, {"char_text": "o", "char_box": [33.0, 2.0, 43.0, 2.0, 36.0, 12.0, 25.0, 12.0]}, {"char_text": "m", "char_box": [46.0, 2.0, 61.0, 2.0, 53.0, 12.0, 39.0, 12.0]}, {"char_text": ":", "char_box": [61.0, 2.0, 69.0, 2.0, 63.0, 12.0, 55.0, 12.0]}], "text": "From:"}
KIE: Difference between CloseSet & OpenSet¶
Being trained on WildReceipt, SDMG-R, or other KIE models, can identify the types of text boxes on a receipt picture. But what SDMG-R can do is far more beyond that. For example, it’s able to identify key-value pairs on the picture. To demonstrate such ability and hopefully facilitate future research, we release a demonstrative version of WildReceiptOpenset annotated in OpenSet format, and provide a full training/testing pipeline for KIE models such as SDMG-R. Since it might be a confusing update, we’ll elaborate on the key differences between the OpenSet and CloseSet format, taking WildReceipt as an example.
CloseSet¶
WildReceipt (“CloseSet”) divides text boxes into 26 categories. There are 12 key-value pairs of fine-grained key information categories, such as (Prod_item_value
, Prod_item_key
), (Prod_price_value
, Prod_price_key
) and (Tax_value
, Tax_key
), plus two more “do not care” categories: Ignore
and Others
.
The objective of CloseSet SDMGR is to predict which category fits the text box best, but it will not predict the relations among text boxes. For instance, if there are four text boxes “Hamburger”, “Hotdog”, “$1” and “$2” on the receipt, the model may assign Prod_item_value
to the first two boxes and Prod_price_value
to the last two, but it can’t tell if Hamburger sells for $1 or $2. However, this could be achieved in the open-set variant.

Warning
A *_key
and *_value
pair do not necessarily have to both appear on the receipt. For example, we usually won’t see Prod_item_key
appearing on the receipt, while there can be multiple boxes annotated as Pred_item_value
. In contrast, Tax_key
and Tax_value
are likely to appear together since they’re usually structured as Tax
: 11.02
on the receipt.
OpenSet¶
In OpenSet, all text boxes, or nodes, have only 4 possible categories: background
, key
, value
, and others
. The connectivity between nodes are annotated as edge labels. If a pair of key-value nodes have the same edge label, they are connected by an valid edge.
Multiple nodes can have the same edge label. However, only key and value nodes will be linked by edges. The nodes of same category will never be connected.
When making OpenSet annotations, each node must have an edge label. It should be an unique one if it falls into non-key
non-value
categories.
Note
You can merge background
to others
if telling background apart is not important, and we provide this choice in the conversion script for WildReceipt .
Converting WildReceipt from CloseSet to OpenSet¶
We provide a conversion script that converts WildRecipt-like dataset to OpenSet format. This script links every key
-value
pairs following the rules above. Here’s an example illustration: (For better understanding, all the node labels are presented as texts)
box_content | closeset_node_label | closeset_edge_label | openset_node_label | openset_edge_label |
---|---|---|---|---|
hello | Ignore | - | Others | 0 |
world | Ignore | - | Others | 1 |
Actor | Actor_key | - | Key | 2 |
Tom | Actor_value | - | Value | 2 |
Tony | Actor_value | - | Value | 2 |
Tim | Actor_value | - | Value | 2 |
something | Ignore | - | Others | 3 |
Actress | Actress_key | - | Key | 4 |
Lucy | Actress_value | - | Value | 4 |
Zora | Actress_value | - | Value | 4 |
Warning
A common request from our community is to extract the relations between food items and food prices. In this case, this conversion script is not you need. Wildrecipt doesn’t provide necessary information to recover this relation. For instance, there are four text boxes “Hamburger”, “Hotdog”, “$1” and “$2” on the receipt, and here’s how they actually look like before and after the conversion:
box_content | closeset_node_label | closeset_edge_label | openset_node_label | openset_edge_label |
---|---|---|---|---|
Hamburger | Prod_item_value | - | Value | 0 |
Hotdog | Prod_item_value | - | Value | 0 |
$1 | Prod_price_value | - | Value | 1 |
$2 | Prod_price_value | - | Value | 1 |
So there won’t be any valid edges connecting them. Nevertheless, OpenSet format is far more general than CloseSet, so this task can be achieved by annotating the data from scratch.
box_content | openset_node_label | openset_edge_label |
---|---|---|
Hamburger | Value | 0 |
Hotdog | Value | 1 |
$1 | Value | 0 |
$2 | Value | 1 |
Enable Blank Space Recognition¶
It is noteworthy that the LineStrParser
should NOT be used to parse the annotation files containing multiple blank spaces (in file name or recognition transcriptions). The users have to convert the plain txt
annotations to json lines
to enable space recognition. For example:
% A plain txt annotation file that contains blank spaces
test/img 1.jpg Hello World!
test/img 2.jpg Hello Open MMLab!
test/img 3.jpg Hello MMOCR!
The LineStrParser
will split the above annotation line to pieces (e.g. [‘test/img’, ‘1.jpg’, ‘Hello’, ‘World!’]) that cannot be matched to the keys
(e.g. [‘filename’, ‘text’]). Therefore, we need to convert it to a json line format by json.dumps
(check here to see how to dump jsonl
), and then the annotation file will look like as follows:
% A json line annotation file that contains blank spaces
{"filename": "test/img 1.jpg", "text": "Hello World!"}
{"filename": "test/img 2.jpg", "text": "Hello Open MMLab!"}
{"filename": "test/img 3.jpg", "text": "Hello MMOCR!"}
After converting the annotation format, you just need to set the parser arguments as:
parser=dict(
type='LineJsonParser',
keys=['filename', 'text']))
Besides, you need to specify a dict that contains blank space to enable blank recognition. Particularly, MMOCR provides two built-in dicts DICT37
and DICT91
that contain blank space. For example, change the default dict_type
in configs/_base_/recog_models/crnn.py
to DICT37
.
label_convertor = dict(
type='CTCConvertor', dict_type='DICT37', with_unknown=False, lower=True) # ['DICT36', 'DICT37', 'DICT90', 'DICT91']
Statistics¶
Number of checkpoints: 33
Number of configs: 26
Number of papers: 19
ALGORITHM: 19
Key Information Extraction Models¶
Number of checkpoints: 3
Number of configs: 3
Number of papers: 1
[ALGORITHM] Spatial Dual-Modality Graph Reasoning for Key Information Extraction
Named Entity Recognition Models¶
Number of checkpoints: 1
Number of configs: 1
Number of papers: 1
[ALGORITHM] Bert: Pre-Training of Deep Bidirectional Transformers for Language Understanding
Text Detection Models¶
Number of checkpoints: 15
Number of configs: 11
Number of papers: 8
[ALGORITHM] Deep Relational Reasoning Graph Network for Arbitrary Shape Text Detection
[ALGORITHM] Efficient and Accurate Arbitrary-Shaped Text Detection With Pixel Aggregation Network
[ALGORITHM] Fourier Contour Embedding for Arbitrary-Shaped Text Detection
[ALGORITHM] Mask R-CNN
[ALGORITHM] Real-Time Scene Text Detection With Differentiable Binarization and Adaptive Scale Fusion
[ALGORITHM] Real-Time Scene Text Detection With Differentiable Binarization
[ALGORITHM] Shape Robust Text Detection With Progressive Scale Expansion Network
[ALGORITHM] Textsnake: A Flexible Representation for Detecting Text of Arbitrary Shapes
Text Recognition Models¶
Number of checkpoints: 14
Number of configs: 11
Number of papers: 9
[ALGORITHM] An End-to-End Trainable Neural Network for Image-Based Sequence Recognition and Its Application to Scene Text Recognition
[ALGORITHM] Nrtr: A No-Recurrence Sequence-to-Sequence Model for Scene Text Recognition
[ALGORITHM] On Recognizing Texts of Arbitrary Shapes With 2d Self-Attention
[ALGORITHM] Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition
[ALGORITHM] Robust Scene Text Recognition With Automatic Rectification
[ALGORITHM] Robustscanner: Dynamically Enhancing Positional Clues for Robust Text Recognition
[ALGORITHM] Segocr Simple Baseline.
[ALGORITHM] Show, Attend and Read: A Simple and Strong Baseline for Irregular Text Recognition
[ALGORITHM] {Master
Model Architecture Summary¶
MMOCR has implemented many models that support various tasks. Depending on the type of tasks, these models have different architectural designs and, therefore, might be a bit confusing for beginners to master. We release a primary design doc to clearly illustrate the basic task-specific architectures and provide quick pointers to docstrings of model components to aid users’ understanding.
Text Detection Models¶

The design of text detectors is similar to SingleStageDetector in MMDetection. The feature of an image was first extracted by backbone
(e.g., ResNet), and neck
further processes raw features into a head-ready format, where the models in MMOCR usually adapt the variants of FPN to extract finer-grained multi-level features. bbox_head
is the core of text detectors, and its implementation varies in different models.
When training, the output of bbox_head
is directly fed into the loss
module, which compares the output with the ground truth and generates a loss dictionary for optimizer’s use. When testing, Postprocessor
converts the outputs from bbox_head
to bounding boxes, which will be used for evaluation metrics (e.g., hmean-iou) and visualization.
DBNet¶
Backbone: mmdet.ResNet
Neck: FPNC
Bbox_head: DBHead
Loss: DBLoss
Postprocessor: DBPostprocessor
DRRG¶
Backbone: mmdet.ResNet
Neck: FPN_UNet
Bbox_head: DRRGHead
Loss: DRRGLoss
Postprocessor: DRRGPostprocessor
FCENet¶
Backbone: mmdet.ResNet
Neck: mmdet.FPN
Bbox_head: FCEHead
Loss: FCELoss
Postprocessor: FCEPostprocessor
Mask R-CNN¶
We use the same architecture as in MMDetection. See MMDetection’s config documentation for details.
PANet¶
Backbone: mmdet.ResNet
Neck: FPEM_FFM
Bbox_head: PANHead
Loss: PANLoss
Postprocessor: PANPostprocessor
PSENet¶
Backbone: mmdet.ResNet
Neck: FPNF
Bbox_head: PSEHead
Loss: PSELoss
Postprocessor: PSEPostprocessor
Textsnake¶
Backbone: mmdet.ResNet
Neck: FPN_UNet
Bbox_head: TextSnakeHead
Loss: TextSnakeLoss
Postprocessor: TextSnakePostprocessor
Text Recognition Models¶
Most of the implemented recognizers use the following architecture:

preprocessor
refers to any network that processes images before they are fed to backbone
. encoder
encodes images features into a hidden vector, which is then transcribed into text tokens by decoder
.
The architecture diverges at training and test phases. The loss module returns a dictionary during training. In testing, converter
is invoked to convert raw features into texts, which are wrapped into a dictionary together with confidence scores. Users can access the dictionary with the text
and score
keys to query the recognition result.
ABINet¶
Preprocessor: None
Backbone: ResNetABI
Encoder: ABIVisionModel
Decoder: ABIVisionDecoder
Fuser: ABIFuser
Loss: ABILoss
Converter: ABIConvertor
Note
Fuser fuses the feature output from encoder and decoder before generating the final text outputs and computing the loss in full ABINet.
CRNN¶
Preprocessor: None
Backbone: VeryDeepVgg
Encoder: None
Decoder: CRNNDecoder
Loss: CTCLoss
Converter: CTCConvertor
CRNN with TPS-based STN¶
Preprocessor: TPSPreprocessor
Backbone: VeryDeepVgg
Encoder: None
Decoder: CRNNDecoder
Loss: CTCLoss
Converter: CTCConvertor
MASTER¶
Preprocessor: None
Backbone: ResNet
Encoder: None
Decoder: MasterDecoder
Loss: TFLoss
Converter: AttnConvertor
NRTR¶
Preprocessor: None
Backbone: ResNet31OCR
Encoder: NRTREncoder
Decoder: NRTRDecoder
Loss: TFLoss
Converter: AttnConvertor
RobustScanner¶
Preprocessor: None
Backbone: ResNet31OCR
Encoder: ChannelReductionEncoder
Decoder: ChannelReductionEncoder
Loss: SARLoss
Converter: AttnConvertor
SAR¶
Preprocessor: None
Backbone: ResNet31OCR
Encoder: SAREncoder
Decoder: ParallelSARDecoder
Loss: SARLoss
Converter: AttnConvertor
SATRN¶
Preprocessor: None
Backbone: ShallowCNN
Encoder: SatrnEncoder
Decoder: NRTRDecoder
Loss: TFLoss
Converter: AttnConvertor
SegOCR¶
Backbone: ResNet31OCR
Neck: FPNOCR
Head: SegHead
Loss: SegLoss
Converter: SegConvertor
Note
SegOCR’s architecture is an exception - it is closer to text detection models.
Key Information Extraction Models¶

The architecture of key information extraction (KIE) models is similar to text detection models, except for the extra feature extractor. As a downstream task of OCR, KIE models are required to run with bounding box annotations indicating the locations of text instances, from which an ROI extractor extracts the cropped features for bbox_head
to discover relations among them.
The output containing edges and nodes information from bbox_head
is sufficient for test and inference. Computation of loss also relies on such information.
SDMGR¶
Backbone: UNet
Neck: None
Extractor: mmdet.SingleRoIExtractor
Bbox_head: SDMGRHead
Loss: SDMGRLoss
Text Detection Models¶
DBNet¶
Real-time Scene Text Detection with Differentiable Binarization
Abstract¶
Recently, segmentation-based methods are quite popular in scene text detection, as the segmentation results can more accurately describe scene text of various shapes such as curve text. However, the post-processing of binarization is essential for segmentation-based detection, which converts probability maps produced by a segmentation method into bounding boxes/regions of text. In this paper, we propose a module named Differentiable Binarization (DB), which can perform the binarization process in a segmentation network. Optimized along with a DB module, a segmentation network can adaptively set the thresholds for binarization, which not only simplifies the post-processing but also enhances the performance of text detection. Based on a simple segmentation network, we validate the performance improvements of DB on five benchmark datasets, which consistently achieves state-of-the-art results, in terms of both detection accuracy and speed. In particular, with a light-weight backbone, the performance improvements by DB are significant so that we can look for an ideal tradeoff between detection accuracy and efficiency. Specifically, with a backbone of ResNet-18, our detector achieves an F-measure of 82.8, running at 62 FPS, on the MSRA-TD500 dataset.

Results and models¶
Citation¶
@article{Liao_Wan_Yao_Chen_Bai_2020,
title={Real-Time Scene Text Detection with Differentiable Binarization},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
author={Liao, Minghui and Wan, Zhaoyi and Yao, Cong and Chen, Kai and Bai, Xiang},
year={2020},
pages={11474-11481}}
DBNetpp¶
Real-Time Scene Text Detection with Differentiable Binarization and Adaptive Scale Fusion
Abstract¶
Recently, segmentation-based scene text detection methods have drawn extensive attention in the scene text detection field, because of their superiority in detecting the text instances of arbitrary shapes and extreme aspect ratios, profiting from the pixel-level descriptions. However, the vast majority of the existing segmentation-based approaches are limited to their complex post-processing algorithms and the scale robustness of their segmentation models, where the post-processing algorithms are not only isolated to the model optimization but also time-consuming and the scale robustness is usually strengthened by fusing multi-scale feature maps directly. In this paper, we propose a Differentiable Binarization (DB) module that integrates the binarization process, one of the most important steps in the post-processing procedure, into a segmentation network. Optimized along with the proposed DB module, the segmentation network can produce more accurate results, which enhances the accuracy of text detection with a simple pipeline. Furthermore, an efficient Adaptive Scale Fusion (ASF) module is proposed to improve the scale robustness by fusing features of different scales adaptively. By incorporating the proposed DB and ASF with the segmentation network, our proposed scene text detector consistently achieves state-of-the-art results, in terms of both detection accuracy and speed, on five standard benchmarks.

Results and models¶
Citation¶
@article{liao2022real,
title={Real-Time Scene Text Detection with Differentiable Binarization and Adaptive Scale Fusion},
author={Liao, Minghui and Zou, Zhisheng and Wan, Zhaoyi and Yao, Cong and Bai, Xiang},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
year={2022},
publisher={IEEE}
}
DRRG¶
Deep relational reasoning graph network for arbitrary shape text detection
Abstract¶
Arbitrary shape text detection is a challenging task due to the high variety and complexity of scenes texts. In this paper, we propose a novel unified relational reasoning graph network for arbitrary shape text detection. In our method, an innovative local graph bridges a text proposal model via Convolutional Neural Network (CNN) and a deep relational reasoning network via Graph Convolutional Network (GCN), making our network end-to-end trainable. To be concrete, every text instance will be divided into a series of small rectangular components, and the geometry attributes (e.g., height, width, and orientation) of the small components will be estimated by our text proposal model. Given the geometry attributes, the local graph construction model can roughly establish linkages between different text components. For further reasoning and deducing the likelihood of linkages between the component and its neighbors, we adopt a graph-based network to perform deep relational reasoning on local graphs. Experiments on public available datasets demonstrate the state-of-the-art performance of our method.

Results and models¶
CTW1500¶
Method | Pretrained Model | Training set | Test set | ##epochs | Test size | Recall | Precision | Hmean | Download |
---|---|---|---|---|---|---|---|---|---|
DRRG | ImageNet | CTW1500 Train | CTW1500 Test | 1200 | 640 | 0.822 (0.791) | 0.858 (0.862) | 0.840 (0.825) | model \ log |
Note
We’ve upgraded our IoU backend from Polygon3
to shapely
. There are some performance differences for some models due to the backends’ different logics to handle invalid polygons (more info here). New evaluation result is presented in brackets and new logs will be uploaded soon.
Citation¶
@article{zhang2020drrg,
title={Deep relational reasoning graph network for arbitrary shape text detection},
author={Zhang, Shi-Xue and Zhu, Xiaobin and Hou, Jie-Bo and Liu, Chang and Yang, Chun and Wang, Hongfa and Yin, Xu-Cheng},
booktitle={CVPR},
pages={9699-9708},
year={2020}
}
FCENet¶
Fourier Contour Embedding for Arbitrary-Shaped Text Detection
Abstract¶
One of the main challenges for arbitrary-shaped text detection is to design a good text instance representation that allows networks to learn diverse text geometry variances. Most of existing methods model text instances in image spatial domain via masks or contour point sequences in the Cartesian or the polar coordinate system. However, the mask representation might lead to expensive post-processing, while the point sequence one may have limited capability to model texts with highly-curved shapes. To tackle these problems, we model text instances in the Fourier domain and propose one novel Fourier Contour Embedding (FCE) method to represent arbitrary shaped text contours as compact signatures. We further construct FCENet with a backbone, feature pyramid networks (FPN) and a simple post-processing with the Inverse Fourier Transformation (IFT) and Non-Maximum Suppression (NMS). Different from previous methods, FCENet first predicts compact Fourier signatures of text instances, and then reconstructs text contours via IFT and NMS during test. Extensive experiments demonstrate that FCE is accurate and robust to fit contours of scene texts even with highly-curved shapes, and also validate the effectiveness and the good generalization of FCENet for arbitrary-shaped text detection. Furthermore, experimental results show that our FCENet is superior to the state-of-the-art (SOTA) methods on CTW1500 and Total-Text, especially on challenging highly-curved text subset.

Results and models¶
Citation¶
@InProceedings{zhu2021fourier,
title={Fourier Contour Embedding for Arbitrary-Shaped Text Detection},
author={Yiqin Zhu and Jianyong Chen and Lingyu Liang and Zhanghui Kuang and Lianwen Jin and Wayne Zhang},
year={2021},
booktitle = {CVPR}
}
Mask R-CNN¶
Abstract¶
We present a conceptually simple, flexible, and general framework for object instance segmentation. Our approach efficiently detects objects in an image while simultaneously generating a high-quality segmentation mask for each instance. The method, called Mask R-CNN, extends Faster R-CNN by adding a branch for predicting an object mask in parallel with the existing branch for bounding box recognition. Mask R-CNN is simple to train and adds only a small overhead to Faster R-CNN, running at 5 fps. Moreover, Mask R-CNN is easy to generalize to other tasks, e.g., allowing us to estimate human poses in the same framework. We show top results in all three tracks of the COCO suite of challenges, including instance segmentation, bounding-box object detection, and person keypoint detection. Without bells and whistles, Mask R-CNN outperforms all existing, single-model entries on every task, including the COCO 2016 challenge winners. We hope our simple and effective approach will serve as a solid baseline and help ease future research in instance-level recognition.

Results and models¶
CTW1500¶
Method | Pretrained Model | Training set | Test set | ##epochs | Test size | Recall | Precision | Hmean | Download |
---|---|---|---|---|---|---|---|---|---|
MaskRCNN | ImageNet | CTW1500 Train | CTW1500 Test | 160 | 1600 | 0.753 | 0.712 | 0.732 | model | log |
ICDAR2015¶
Method | Pretrained Model | Training set | Test set | ##epochs | Test size | Recall | Precision | Hmean | Download |
---|---|---|---|---|---|---|---|---|---|
MaskRCNN | ImageNet | ICDAR2015 Train | ICDAR2015 Test | 160 | 1920 | 0.783 | 0.872 | 0.825 | model | log |
ICDAR2017¶
Method | Pretrained Model | Training set | Test set | ##epochs | Test size | Recall | Precision | Hmean | Download |
---|---|---|---|---|---|---|---|---|---|
MaskRCNN | ImageNet | ICDAR2017 Train | ICDAR2017 Val | 160 | 1600 | 0.754 | 0.827 | 0.789 | model | log |
Note
We tuned parameters with the techniques in Pyramid Mask Text Detector
Citation¶
@INPROCEEDINGS{8237584,
author={K. {He} and G. {Gkioxari} and P. {Dollár} and R. {Girshick}},
booktitle={2017 IEEE International Conference on Computer Vision (ICCV)},
title={Mask R-CNN},
year={2017},
pages={2980-2988},
doi={10.1109/ICCV.2017.322}}
PANet¶
Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network
Abstract¶
Scene text detection, an important step of scene text reading systems, has witnessed rapid development with convolutional neural networks. Nonetheless, two main challenges still exist and hamper its deployment to real-world applications. The first problem is the trade-off between speed and accuracy. The second one is to model the arbitrary-shaped text instance. Recently, some methods have been proposed to tackle arbitrary-shaped text detection, but they rarely take the speed of the entire pipeline into consideration, which may fall short in practical this http URL this paper, we propose an efficient and accurate arbitrary-shaped text detector, termed Pixel Aggregation Network (PAN), which is equipped with a low computational-cost segmentation head and a learnable post-processing. More specifically, the segmentation head is made up of Feature Pyramid Enhancement Module (FPEM) and Feature Fusion Module (FFM). FPEM is a cascadable U-shaped module, which can introduce multi-level information to guide the better segmentation. FFM can gather the features given by the FPEMs of different depths into a final feature for segmentation. The learnable post-processing is implemented by Pixel Aggregation (PA), which can precisely aggregate text pixels by predicted similarity vectors. Experiments on several standard benchmarks validate the superiority of the proposed PAN. It is worth noting that our method can achieve a competitive F-measure of 79.9% at 84.2 FPS on CTW1500.

Results and models¶
CTW1500¶
Method | Pretrained Model | Training set | Test set | ##epochs | Test size | Recall | Precision | Hmean | Download |
---|---|---|---|---|---|---|---|---|---|
PANet | ImageNet | CTW1500 Train | CTW1500 Test | 600 | 640 | 0.776 (0.717) | 0.838 (0.835) | 0.806 (0.801) | model | log |
ICDAR2015¶
Method | Pretrained Model | Training set | Test set | ##epochs | Test size | Recall | Precision | Hmean | Download |
---|---|---|---|---|---|---|---|---|---|
PANet | ImageNet | ICDAR2015 Train | ICDAR2015 Test | 600 | 736 | 0.734 (0.74) | 0.856 (0.86) | 0.791 (0.795) | model | log |
Note
We’ve upgraded our IoU backend from Polygon3
to shapely
. There are some performance differences for some models due to the backends’ different logics to handle invalid polygons (more info here). New evaluation result is presented in brackets and new logs will be uploaded soon.
Citation¶
@inproceedings{WangXSZWLYS19,
author={Wenhai Wang and Enze Xie and Xiaoge Song and Yuhang Zang and Wenjia Wang and Tong Lu and Gang Yu and Chunhua Shen},
title={Efficient and Accurate Arbitrary-Shaped Text Detection With Pixel Aggregation Network},
booktitle={ICCV},
pages={8439--8448},
year={2019}
}
PSENet¶
Shape robust text detection with progressive scale expansion network
Abstract¶
Scene text detection has witnessed rapid progress especially with the recent development of convolutional neural networks. However, there still exists two challenges which prevent the algorithm into industry applications. On the one hand, most of the state-of-art algorithms require quadrangle bounding box which is in-accurate to locate the texts with arbitrary shape. On the other hand, two text instances which are close to each other may lead to a false detection which covers both instances. Traditionally, the segmentation-based approach can relieve the first problem but usually fail to solve the second challenge. To address these two challenges, in this paper, we propose a novel Progressive Scale Expansion Network (PSENet), which can precisely detect text instances with arbitrary shapes. More specifically, PSENet generates the different scale of kernels for each text instance, and gradually expands the minimal scale kernel to the text instance with the complete shape. Due to the fact that there are large geometrical margins among the minimal scale kernels, our method is effective to split the close text instances, making it easier to use segmentation-based methods to detect arbitrary-shaped text instances. Extensive experiments on CTW1500, Total-Text, ICDAR 2015 and ICDAR 2017 MLT validate the effectiveness of PSENet. Notably, on CTW1500, a dataset full of long curve texts, PSENet achieves a F-measure of 74.3% at 27 FPS, and our best F-measure (82.2%) outperforms state-of-art algorithms by 6.6%. The code will be released in the future.

Results and models¶
CTW1500¶
Method | Backbone | Extra Data | Training set | Test set | ##epochs | Test size | Recall | Precision | Hmean | Download |
---|---|---|---|---|---|---|---|---|---|---|
PSENet-4s | ResNet50 | - | CTW1500 Train | CTW1500 Test | 600 | 1280 | 0.728 (0.717) | 0.849 (0.852) | 0.784 (0.779) | model | log |
ICDAR2015¶
Method | Backbone | Extra Data | Training set | Test set | ##epochs | Test size | Recall | Precision | Hmean | Download |
---|---|---|---|---|---|---|---|---|---|---|
PSENet-4s | ResNet50 | - | IC15 Train | IC15 Test | 600 | 2240 | 0.784 (0.753) | 0.831 (0.867) | 0.807 (0.806) | model | log |
PSENet-4s | ResNet50 | pretrain on IC17 MLT model | IC15 Train | IC15 Test | 600 | 2240 | 0.834 | 0.861 | 0.847 | model | log |
Note
We’ve upgraded our IoU backend from Polygon3
to shapely
. There are some performance differences for some models due to the backends’ different logics to handle invalid polygons (more info here). New evaluation result is presented in brackets and new logs will be uploaded soon.
Citation¶
@inproceedings{wang2019shape,
title={Shape robust text detection with progressive scale expansion network},
author={Wang, Wenhai and Xie, Enze and Li, Xiang and Hou, Wenbo and Lu, Tong and Yu, Gang and Shao, Shuai},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={9336--9345},
year={2019}
}
Textsnake¶
TextSnake: A Flexible Representation for Detecting Text of Arbitrary Shapes
Abstract¶
Driven by deep neural networks and large scale datasets, scene text detection methods have progressed substantially over the past years, continuously refreshing the performance records on various standard benchmarks. However, limited by the representations (axis-aligned rectangles, rotated rectangles or quadrangles) adopted to describe text, existing methods may fall short when dealing with much more free-form text instances, such as curved text, which are actually very common in real-world scenarios. To tackle this problem, we propose a more flexible representation for scene text, termed as TextSnake, which is able to effectively represent text instances in horizontal, oriented and curved forms. In TextSnake, a text instance is described as a sequence of ordered, overlapping disks centered at symmetric axes, each of which is associated with potentially variable radius and orientation. Such geometry attributes are estimated via a Fully Convolutional Network (FCN) model. In experiments, the text detector based on TextSnake achieves state-of-the-art or comparable performance on Total-Text and SCUT-CTW1500, the two newly published benchmarks with special emphasis on curved text in natural images, as well as the widely-used datasets ICDAR 2015 and MSRA-TD500. Specifically, TextSnake outperforms the baseline on Total-Text by more than 40% in F-measure.

Results and models¶
Citation¶
@article{long2018textsnake,
title={TextSnake: A Flexible Representation for Detecting Text of Arbitrary Shapes},
author={Long, Shangbang and Ruan, Jiaqiang and Zhang, Wenjie and He, Xin and Wu, Wenhao and Yao, Cong},
booktitle={ECCV},
pages={20-36},
year={2018}
}
Text Recognition Models¶
ABINet¶
Abstract¶
Linguistic knowledge is of great benefit to scene text recognition. However, how to effectively model linguistic rules in end-to-end deep networks remains a research challenge. In this paper, we argue that the limited capacity of language models comes from: 1) implicitly language modeling; 2) unidirectional feature representation; and 3) language model with noise input. Correspondingly, we propose an autonomous, bidirectional and iterative ABINet for scene text recognition. Firstly, the autonomous suggests to block gradient flow between vision and language models to enforce explicitly language modeling. Secondly, a novel bidirectional cloze network (BCN) as the language model is proposed based on bidirectional feature representation. Thirdly, we propose an execution manner of iterative correction for language model which can effectively alleviate the impact of noise input. Additionally, based on the ensemble of iterative predictions, we propose a self-training method which can learn from unlabeled images effectively. Extensive experiments indicate that ABINet has superiority on low-quality images and achieves state-of-the-art results on several mainstream benchmarks. Besides, the ABINet trained with ensemble self-training shows promising improvement in realizing human-level recognition.

Dataset¶
Train Dataset¶
trainset | instance_num | repeat_num | note |
---|---|---|---|
Syn90k | 8919273 | 1 | synth |
SynthText | 7239272 | 1 | alphanumeric |
Test Dataset¶
testset | instance_num | note |
---|---|---|
IIIT5K | 3000 | regular |
SVT | 647 | regular |
IC13 | 1015 | regular |
IC15 | 2077 | irregular |
SVTP | 645 | irregular |
CT80 | 288 | irregular |
Results and models¶
methods | pretrained | Regular Text | Irregular Text | download | ||||
---|---|---|---|---|---|---|---|---|
IIIT5K | SVT | IC13 | IC15 | SVTP | CT80 | |||
ABINet-Vision | - | 94.7 | 91.7 | 93.6 | 83.0 | 85.1 | 86.5 | model | log |
ABINet | Pretrained | 95.7 | 94.6 | 95.7 | 85.1 | 90.4 | 90.3 | model | log1 | log2 |
Note
ABINet allows its encoder to run and be trained without decoder and fuser. Its encoder is designed to recognize texts as a stand-alone model and therefore can work as an independent text recognizer. We release it as ABINet-Vision.
Facts about the pretrained model: MMOCR does not have a systematic pipeline to pretrain the language model (LM) yet, thus the weights of LM are converted from the official pretrained model. The weights of ABINet-Vision are directly used as the vision model of ABINet.
Due to some technical issues, the training process of ABINet was interrupted at the 13th epoch and we resumed it later. Both logs are released for full reference.
The model architecture in the logs looks slightly different from the final released version, since it was refactored afterward. However, both architectures are essentially equivalent.
Citation¶
@article{fang2021read,
title={Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition},
author={Fang, Shancheng and Xie, Hongtao and Wang, Yuxin and Mao, Zhendong and Zhang, Yongdong},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2021}
}
CRNN¶
Abstract¶
Image-based sequence recognition has been a long-standing research topic in computer vision. In this paper, we investigate the problem of scene text recognition, which is among the most important and challenging tasks in image-based sequence recognition. A novel neural network architecture, which integrates feature extraction, sequence modeling and transcription into a unified framework, is proposed. Compared with previous systems for scene text recognition, the proposed architecture possesses four distinctive properties: (1) It is end-to-end trainable, in contrast to most of the existing algorithms whose components are separately trained and tuned. (2) It naturally handles sequences in arbitrary lengths, involving no character segmentation or horizontal scale normalization. (3) It is not confined to any predefined lexicon and achieves remarkable performances in both lexicon-free and lexicon-based scene text recognition tasks. (4) It generates an effective yet much smaller model, which is more practical for real-world application scenarios. The experiments on standard benchmarks, including the IIIT-5K, Street View Text and ICDAR datasets, demonstrate the superiority of the proposed algorithm over the prior arts. Moreover, the proposed algorithm performs well in the task of image-based music score recognition, which evidently verifies the generality of it.

Dataset¶
Train Dataset¶
trainset | instance_num | repeat_num | note |
---|---|---|---|
Syn90k | 8919273 | 1 | synth |
Test Dataset¶
testset | instance_num | note |
---|---|---|
IIIT5K | 3000 | regular |
SVT | 647 | regular |
IC13 | 1015 | regular |
IC15 | 2077 | irregular |
SVTP | 645 | irregular |
CT80 | 288 | irregular |
Results and models¶
methods | Regular Text | Irregular Text | download | |||||
---|---|---|---|---|---|---|---|---|
methods | IIIT5K | SVT | IC13 | IC15 | SVTP | CT80 | ||
CRNN | 80.5 | 81.5 | 86.5 | 54.1 | 59.1 | 55.6 | model | log |
Citation¶
@article{shi2016end,
title={An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition},
author={Shi, Baoguang and Bai, Xiang and Yao, Cong},
journal={IEEE transactions on pattern analysis and machine intelligence},
year={2016}
}
MASTER¶
MASTER: Multi-aspect non-local network for scene text recognition
Abstract¶
Attention-based scene text recognizers have gained huge success, which leverages a more compact intermediate representation to learn 1d- or 2d- attention by a RNN-based encoder-decoder architecture. However, such methods suffer from attention-drift problem because high similarity among encoded features leads to attention confusion under the RNN-based local attention mechanism. Moreover, RNN-based methods have low efficiency due to poor parallelization. To overcome these problems, we propose the MASTER, a self-attention based scene text recognizer that (1) not only encodes the input-output attention but also learns self-attention which encodes feature-feature and target-target relationships inside the encoder and decoder and (2) learns a more powerful and robust intermediate representation to spatial distortion, and (3) owns a great training efficiency because of high training parallelization and a high-speed inference because of an efficient memory-cache mechanism. Extensive experiments on various benchmarks demonstrate the superior performance of our MASTER on both regular and irregular scene text.
Dataset¶
Train Dataset¶
trainset | instance_num | repeat_num | source |
---|---|---|---|
SynthText | 7266686 | 1 | synth |
SynthAdd | 1216889 | 1 | synth |
Syn90k | 8919273 | 1 | synth |
Test Dataset¶
testset | instance_num | type |
---|---|---|
IIIT5K | 3000 | regular |
SVT | 647 | regular |
IC13 | 1015 | regular |
IC15 | 2077 | irregular |
SVTP | 645 | irregular |
CT80 | 288 | irregular |
Results and Models¶
Methods | Backbone | Regular Text | Irregular Text | download | |||||
---|---|---|---|---|---|---|---|---|---|
IIIT5K | SVT | IC13 | IC15 | SVTP | CT80 | ||||
MASTER | R31-GCAModule | 95.27 | 89.8 | 95.17 | 77.03 | 82.95 | 89.93 | model | log |
Citation¶
@article{Lu2021MASTER,
title={{MASTER}: Multi-Aspect Non-local Network for Scene Text Recognition},
author={Ning Lu and Wenwen Yu and Xianbiao Qi and Yihao Chen and Ping Gong and Rong Xiao and Xiang Bai},
journal={Pattern Recognition},
year={2021}
}
NRTR¶
NRTR: A No-Recurrence Sequence-to-Sequence Model For Scene Text Recognition
Abstract¶
Scene text recognition has attracted a great many researches due to its importance to various applications. Existing methods mainly adopt recurrence or convolution based networks. Though have obtained good performance, these methods still suffer from two limitations: slow training speed due to the internal recurrence of RNNs, and high complexity due to stacked convolutional layers for long-term feature extraction. This paper, for the first time, proposes a no-recurrence sequence-to-sequence text recognizer, named NRTR, that dispenses with recurrences and convolutions entirely. NRTR follows the encoder-decoder paradigm, where the encoder uses stacked self-attention to extract image features, and the decoder applies stacked self-attention to recognize texts based on encoder output. NRTR relies solely on self-attention mechanism thus could be trained with more parallelization and less complexity. Considering scene image has large variation in text and background, we further design a modality-transform block to effectively transform 2D input images to 1D sequences, combined with the encoder to extract more discriminative features. NRTR achieves state-of-the-art or highly competitive performance on both regular and irregular benchmarks, while requires only a small fraction of training time compared to the best model from the literature (at least 8 times faster).

Dataset¶
Train Dataset¶
trainset | instance_num | repeat_num | source |
---|---|---|---|
SynthText | 7266686 | 1 | synth |
Syn90k | 8919273 | 1 | synth |
Test Dataset¶
testset | instance_num | type |
---|---|---|
IIIT5K | 3000 | regular |
SVT | 647 | regular |
IC13 | 1015 | regular |
IC15 | 2077 | irregular |
SVTP | 645 | irregular |
CT80 | 288 | irregular |
Results and Models¶
Methods | Backbone | Regular Text | Irregular Text | download | |||||
---|---|---|---|---|---|---|---|---|---|
IIIT5K | SVT | IC13 | IC15 | SVTP | CT80 | ||||
NRTR | R31-1/16-1/8 | 94.7 | 87.3 | 94.3 | 73.5 | 78.9 | 85.1 | model | log | |
NRTR | R31-1/8-1/4 | 95.2 | 90.0 | 94.0 | 74.1 | 79.4 | 88.2 | model | log |
Note
For backbone
R31-1/16-1/8
:The output consists of 92 classes, including 26 lowercase letters, 26 uppercase letters, 28 symbols, 10 digital numbers, 1 unknown token and 1 end-of-sequence token.
The encoder-block number is 6.
1/16-1/8
means the height of feature from backbone is 1/16 of input image, where 1/8 for width.
For backbone
R31-1/8-1/4
:The output consists of 92 classes, including 26 lowercase letters, 26 uppercase letters, 28 symbols, 10 digital numbers, 1 unknown token and 1 end-of-sequence token.
The encoder-block number is 6.
1/8-1/4
means the height of feature from backbone is 1/8 of input image, where 1/4 for width.
Citation¶
@inproceedings{sheng2019nrtr,
title={NRTR: A no-recurrence sequence-to-sequence model for scene text recognition},
author={Sheng, Fenfen and Chen, Zhineng and Xu, Bo},
booktitle={2019 International Conference on Document Analysis and Recognition (ICDAR)},
pages={781--786},
year={2019},
organization={IEEE}
}
RobustScanner¶
RobustScanner: Dynamically Enhancing Positional Clues for Robust Text Recognition
Abstract¶
The attention-based encoder-decoder framework has recently achieved impressive results for scene text recognition, and many variants have emerged with improvements in recognition quality. However, it performs poorly on contextless texts (e.g., random character sequences) which is unacceptable in most of real application scenarios. In this paper, we first deeply investigate the decoding process of the decoder. We empirically find that a representative character-level sequence decoder utilizes not only context information but also positional information. Contextual information, which the existing approaches heavily rely on, causes the problem of attention drift. To suppress such side-effect, we propose a novel position enhancement branch, and dynamically fuse its outputs with those of the decoder attention module for scene text recognition. Specifically, it contains a position aware module to enable the encoder to output feature vectors encoding their own spatial positions, and an attention module to estimate glimpses using the positional clue (i.e., the current decoding time step) only. The dynamic fusion is conducted for more robust feature via an element-wise gate mechanism. Theoretically, our proposed method, dubbed \emph{RobustScanner}, decodes individual characters with dynamic ratio between context and positional clues, and utilizes more positional ones when the decoding sequences with scarce context, and thus is robust and practical. Empirically, it has achieved new state-of-the-art results on popular regular and irregular text recognition benchmarks while without much performance drop on contextless benchmarks, validating its robustness in both contextual and contextless application scenarios.

Dataset¶
Results and Models¶
Methods | GPUs | Regular Text | Irregular Text | download | |||||
---|---|---|---|---|---|---|---|---|---|
IIIT5K | SVT | IC13 | IC15 | SVTP | CT80 | ||||
RobustScanner | 16 | 95.1 | 89.2 | 93.1 | 77.8 | 80.3 | 90.3 | model | log |
References¶
[1] Li, Hui and Wang, Peng and Shen, Chunhua and Zhang, Guyu. Show, attend and read: A simple and strong baseline for irregular text recognition. In AAAI 2019.
Citation¶
@inproceedings{yue2020robustscanner,
title={RobustScanner: Dynamically Enhancing Positional Clues for Robust Text Recognition},
author={Yue, Xiaoyu and Kuang, Zhanghui and Lin, Chenhao and Sun, Hongbin and Zhang, Wayne},
booktitle={European Conference on Computer Vision},
year={2020}
}
SAR¶
Show, Attend and Read: A Simple and Strong Baseline for Irregular Text Recognition
Abstract¶
Recognizing irregular text in natural scene images is challenging due to the large variance in text appearance, such as curvature, orientation and distortion. Most existing approaches rely heavily on sophisticated model designs and/or extra fine-grained annotations, which, to some extent, increase the difficulty in algorithm implementation and data collection. In this work, we propose an easy-to-implement strong baseline for irregular scene text recognition, using off-the-shelf neural network components and only word-level annotations. It is composed of a 31-layer ResNet, an LSTM-based encoder-decoder framework and a 2-dimensional attention module. Despite its simplicity, the proposed method is robust and achieves state-of-the-art performance on both regular and irregular scene text recognition benchmarks.

Dataset¶
Results and Models¶
Methods | Backbone | Decoder | Regular Text | Irregular Text | download | |||||
---|---|---|---|---|---|---|---|---|---|---|
IIIT5K | SVT | IC13 | IC15 | SVTP | CT80 | |||||
SAR | R31-1/8-1/4 | ParallelSARDecoder | 95.0 | 89.6 | 93.7 | 79.0 | 82.2 | 88.9 | model | log | |
SAR | R31-1/8-1/4 | SequentialSARDecoder | 95.2 | 88.7 | 92.4 | 78.2 | 81.9 | 89.6 | model | log |
Chinese Dataset¶
Results and Models¶
Methods | Backbone | Decoder | download | |
---|---|---|---|---|
SAR | R31-1/8-1/4 | ParallelSARDecoder | model | log | dict |
Note
R31-1/8-1/4
means the height of feature from backbone is 1/8 of input image, where 1/4 for width.We did not use beam search during decoding.
We implemented two kinds of decoder. Namely,
ParallelSARDecoder
andSequentialSARDecoder
.ParallelSARDecoder
: Parallel decoding during training withLSTM
layer. It would be faster.SequentialSARDecoder
: Sequential Decoding during training withLSTMCell
. It would be easier to understand.
For train dataset.
We did not construct distinct data groups (20 groups in [1]) to train the model group-by-group since it would render model training too complicated.
Instead, we randomly selected
2.4m
patches fromSyn90k
,2.4m
fromSynthText
and1.2m
fromSynthAdd
, and grouped all data together. See config for details.
We used 48 GPUs with
total_batch_size = 64 * 48
in the experiment above to speedup training, while keeping theinitial lr = 1e-3
unchanged.
Citation¶
@inproceedings{li2019show,
title={Show, attend and read: A simple and strong baseline for irregular text recognition},
author={Li, Hui and Wang, Peng and Shen, Chunhua and Zhang, Guyu},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={33},
number={01},
pages={8610--8617},
year={2019}
}
SATRN¶
On Recognizing Texts of Arbitrary Shapes with 2D Self-Attention
Abstract¶
Scene text recognition (STR) is the task of recognizing character sequences in natural scenes. While there have been great advances in STR methods, current methods still fail to recognize texts in arbitrary shapes, such as heavily curved or rotated texts, which are abundant in daily life (e.g. restaurant signs, product labels, company logos, etc). This paper introduces a novel architecture to recognizing texts of arbitrary shapes, named Self-Attention Text Recognition Network (SATRN), which is inspired by the Transformer. SATRN utilizes the self-attention mechanism to describe two-dimensional (2D) spatial dependencies of characters in a scene text image. Exploiting the full-graph propagation of self-attention, SATRN can recognize texts with arbitrary arrangements and large inter-character spacing. As a result, SATRN outperforms existing STR models by a large margin of 5.7 pp on average in “irregular text” benchmarks. We provide empirical analyses that illustrate the inner mechanisms and the extent to which the model is applicable (e.g. rotated and multi-line text). We will open-source the code.

Dataset¶
Train Dataset¶
trainset | instance_num | repeat_num | source |
---|---|---|---|
SynthText | 7266686 | 1 | synth |
Syn90k | 8919273 | 1 | synth |
Test Dataset¶
testset | instance_num | type |
---|---|---|
IIIT5K | 3000 | regular |
SVT | 647 | regular |
IC13 | 1015 | regular |
IC15 | 2077 | irregular |
SVTP | 645 | irregular |
CT80 | 288 | irregular |
Results and Models¶
Methods | Regular Text | Irregular Text | download | |||||
---|---|---|---|---|---|---|---|---|
IIIT5K | SVT | IC13 | IC15 | SVTP | CT80 | |||
Satrn | 96.1 | 93.5 | 95.7 | 84.1 | 88.5 | 90.3 | model | log | |
Satrn_small | 94.7 | 91.3 | 95.4 | 81.9 | 85.9 | 86.5 | model | log |
Citation¶
@article{junyeop2019recognizing,
title={On Recognizing Texts of Arbitrary Shapes with 2D Self-Attention},
author={Junyeop Lee, Sungrae Park, Jeonghun Baek, Seong Joon Oh, Seonghyeon Kim, Hwalsuk Lee},
year={2019}
}
SegOCR¶
Abstract¶
Just a simple Seg-based baseline for text recognition tasks.
Dataset¶
Train Dataset¶
trainset | instance_num | repeat_num | source |
---|---|---|---|
SynthText | 7266686 | 1 | synth |
Test Dataset¶
testset | instance_num | type |
---|---|---|
IIIT5K | 3000 | regular |
SVT | 647 | regular |
IC13 | 1015 | regular |
CT80 | 288 | irregular |
Results and Models¶
Backbone | Neck | Head | Regular Text | Irregular Text | download | ||||
---|---|---|---|---|---|---|---|---|---|
IIIT5K | SVT | IC13 | CT80 | ||||||
R31-1/16 | FPNOCR | 1x | 90.9 | 81.8 | 90.7 | 80.9 | model | log |
Note
R31-1/16
means the size (both height and width ) of feature from backbone is 1/16 of input image.1x
means the size (both height and width) of feature from head is the same with input image.
Citation¶
@unpublished{key,
title={SegOCR Simple Baseline.},
author={},
note={Unpublished Manuscript},
year={2021}
}
CRNN-STN¶
Abstract¶
Image-based sequence recognition has been a long-standing research topic in computer vision. In this paper, we investigate the problem of scene text recognition, which is among the most important and challenging tasks in image-based sequence recognition. A novel neural network architecture, which integrates feature extraction, sequence modeling and transcription into a unified framework, is proposed. Compared with previous systems for scene text recognition, the proposed architecture possesses four distinctive properties: (1) It is end-to-end trainable, in contrast to most of the existing algorithms whose components are separately trained and tuned. (2) It naturally handles sequences in arbitrary lengths, involving no character segmentation or horizontal scale normalization. (3) It is not confined to any predefined lexicon and achieves remarkable performances in both lexicon-free and lexicon-based scene text recognition tasks. (4) It generates an effective yet much smaller model, which is more practical for real-world application scenarios. The experiments on standard benchmarks, including the IIIT-5K, Street View Text and ICDAR datasets, demonstrate the superiority of the proposed algorithm over the prior arts. Moreover, the proposed algorithm performs well in the task of image-based music score recognition, which evidently verifies the generality of it.

Note
We use STN from this paper as the preprocessor and CRNN as the recognition network.
Dataset¶
Train Dataset¶
trainset | instance_num | repeat_num | note |
---|---|---|---|
Syn90k | 8919273 | 1 | synth |
Test Dataset¶
testset | instance_num | note |
---|---|---|
IIIT5K | 3000 | regular |
SVT | 647 | regular |
IC13 | 1015 | regular |
IC15 | 2077 | irregular |
SVTP | 645 | irregular |
CT80 | 288 | irregular |
Results and models¶
methods | Regular Text | Irregular Text | download | |||||
---|---|---|---|---|---|---|---|---|
IIIT5K | SVT | IC13 | IC15 | SVTP | CT80 | |||
CRNN-STN | 80.8 | 81.3 | 85.0 | 59.6 | 68.1 | 53.8 | model | log |
Citation¶
@article{shi2016robust,
title={Robust Scene Text Recognition with Automatic Rectification},
author={Shi, Baoguang and Wang, Xinggang and Lyu, Pengyuan and Yao,
Cong and Bai, Xiang},
year={2016}
}
Key Information Extraction Models¶
SDMGR¶
Spatial Dual-Modality Graph Reasoning for Key Information Extraction
Abstract¶
Key information extraction from document images is of paramount importance in office automation. Conventional template matching based approaches fail to generalize well to document images of unseen templates, and are not robust against text recognition errors. In this paper, we propose an end-to-end Spatial Dual-Modality Graph Reasoning method (SDMG-R) to extract key information from unstructured document images. We model document images as dual-modality graphs, nodes of which encode both the visual and textual features of detected text regions, and edges of which represent the spatial relations between neighboring text regions. The key information extraction is solved by iteratively propagating messages along graph edges and reasoning the categories of graph nodes. In order to roundly evaluate our proposed method as well as boost the future research, we release a new dataset named WildReceipt, which is collected and annotated tailored for the evaluation of key information extraction from document images of unseen templates in the wild. It contains 25 key information categories, a total of about 69000 text boxes, and is about 2 times larger than the existing public datasets. Extensive experiments validate that all information including visual features, textual features and spatial relations can benefit key information extraction. It has been shown that SDMG-R can effectively extract key information from document images of unseen templates, and obtain new state-of-the-art results on the recent popular benchmark SROIE and our WildReceipt. Our code and dataset will be publicly released.

Results and models¶
WildReceipt¶
Method | Modality | Macro F1-Score | Download |
---|---|---|---|
sdmgr_unet16 | Visual + Textual | 0.888 | model | log |
sdmgr_novisual | Textual | 0.871 | model | log |
Note
For
sdmgr_novisual
, images are not needed for training and testing. So fakeimg_prefix
can be used in configs. As well, fakefile_name
can be used in annotation files.
WildReceiptOpenset¶
Method | Modality | Edge F1-Score | Node Macro F1-Score | Node Micro F1-Score | Download |
---|---|---|---|---|---|
sdmgr_novisual | Textual | 0.796 | 0.930 | 0.938 | model | log |
Note
In the case of openset, the number of node categories is unknown or unfixed, and more node category can be added.
To show that our method can handle openset problem, we modify the ground truth of
WildReceipt
toWildReceiptOpenset
. Thenodes
are just classified into 4 classes:background, key, value, others
, while addingedge
labels for each box.The model is used to predict whether two nodes are a pair connecting by a valid edge.
You can learn more about the key differences between CloseSet and OpenSet annotations in our tutorial.
Citation¶
@misc{sun2021spatial,
title={Spatial Dual-Modality Graph Reasoning for Key Information Extraction},
author={Hongbin Sun and Zhanghui Kuang and Xiaoyu Yue and Chenhao Lin and Wayne Zhang},
year={2021},
eprint={2103.14470},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Named Entity Recognition Models¶
Bert¶
Bert: Pre-training of deep bidirectional transformers for language understanding
Abstract¶
We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).

Dataset¶
Train Dataset¶
trainset | text_num | entity_num |
---|---|---|
CLUENER2020 | 10748 | 23338 |
Test Dataset¶
testset | text_num | entity_num |
---|---|---|
CLUENER2020 | 1343 | 2982 |
Results and models¶
Method | Pretrain | Precision | Recall | F1-Score | Download |
---|---|---|---|---|---|
bert_softmax | pretrain | 0.7885 | 0.7998 | 0.7941 | model | log |
Citation¶
@article{devlin2018bert,
title={Bert: Pre-training of deep bidirectional transformers for language understanding},
author={Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},
journal={arXiv preprint arXiv:1810.04805},
year={2018}
}
Text Detection¶
Overview¶
Dataset | Images | Annotation Files | |||
---|---|---|---|---|---|
training | validation | testing | |||
CTW1500 | homepage | - | - | - | |
ICDAR2011 | homepage | - | - | ||
ICDAR2013 | homepage | - | - | - | |
ICDAR2015 | homepage | instances_training.json | - | instances_test.json | |
ICDAR2017 | homepage | instances_training.json | instances_val.json | - | |
Synthtext | homepage | instances_training.lmdb (data.mdb, lock.mdb) | - | - | |
TextOCR | homepage | - | - | - | |
Totaltext | homepage | - | - | - | |
CurvedSynText150k | homepage | Part1 | Part2 | instances_training.json | - | - | |
FUNSD | homepage | - | - | - | |
DeText | homepage | - | - | - | |
NAF | homepage | - | - | - | |
SROIE | homepage | - | - | - | |
Lecture Video DB | homepage | - | - | - | |
LSVT | homepage | - | - | - | |
IMGUR | homepage | - | - | - | |
KAIST | homepage | - | - | - | |
MTWI | homepage | - | - | - | |
COCO Text v2 | homepage | - | - | - | |
ReCTS | homepage | - | - | - | |
IIIT-ILST | homepage | - | - | - | |
VinText | homepage | - | - | - | |
BID | homepage | - | - | - | |
RCTW | homepage | - | - | - | |
HierText | homepage | - | - | - | |
ArT | homepage | - | - | - |
Install AWS CLI (optional)¶
Since there are some datasets that require the AWS CLI to be installed in advance, we provide a quick installation guide here:
curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip" unzip awscliv2.zip sudo ./aws/install ./aws/install -i /usr/local/aws-cli -b /usr/local/bin !aws configure # this command will require you to input keys, you can skip them except # for the Default region name # AWS Access Key ID [None]: # AWS Secret Access Key [None]: # Default region name [None]: us-east-1 # Default output format [None]
Important Note¶
Note
For users who want to train models on CTW1500, ICDAR 2015/2017, and Totaltext dataset, there might be some images containing orientation info in EXIF data. The default OpenCV
backend used in MMCV would read them and apply the rotation on the images. However, their gold annotations are made on the raw pixels, and such
inconsistency results in false examples in the training set. Therefore, users should use dict(type='LoadImageFromFile', color_type='color_ignore_orientation')
in pipelines to change MMCV’s default loading behaviour. (see DBNet’s pipeline config for example)
CTW1500¶
Step0: Read Important Note
Step1: Download
train_images.zip
,test_images.zip
,train_labels.zip
,test_labels.zip
from githubmkdir ctw1500 && cd ctw1500 mkdir imgs && mkdir annotations # For annotations cd annotations wget -O train_labels.zip https://universityofadelaide.box.com/shared/static/jikuazluzyj4lq6umzei7m2ppmt3afyw.zip wget -O test_labels.zip https://cloudstor.aarnet.edu.au/plus/s/uoeFl0pCN9BOCN5/download unzip train_labels.zip && mv ctw1500_train_labels training unzip test_labels.zip -d test cd .. # For images cd imgs wget -O train_images.zip https://universityofadelaide.box.com/shared/static/py5uwlfyyytbb2pxzq9czvu6fuqbjdh8.zip wget -O test_images.zip https://universityofadelaide.box.com/shared/static/t4w48ofnqkdw7jyc4t11nsukoeqk9c3d.zip unzip train_images.zip && mv train_images training unzip test_images.zip && mv test_images test
Step2: Generate
instances_training.json
andinstances_test.json
with following command:python tools/data/textdet/ctw1500_converter.py /path/to/ctw1500 -o /path/to/ctw1500 --split-list training test
The resulting directory structure looks like the following:
├── ctw1500 │ ├── imgs │ ├── annotations │ ├── instances_training.json │ └── instances_val.json
ICDAR 2011 (Born-Digital Images)¶
Step1: Download
Challenge1_Training_Task12_Images.zip
,Challenge1_Training_Task1_GT.zip
,Challenge1_Test_Task12_Images.zip
, andChallenge1_Test_Task1_GT.zip
from homepageTask 1.1: Text Localization (2013 edition)
.mkdir icdar2011 && cd icdar2011 mkdir imgs && mkdir annotations # Download ICDAR 2011 wget https://rrc.cvc.uab.es/downloads/Challenge1_Training_Task12_Images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge1_Training_Task1_GT.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge1_Test_Task12_Images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge1_Test_Task1_GT.zip --no-check-certificate # For images unzip -q Challenge1_Training_Task12_Images.zip -d imgs/training unzip -q Challenge1_Test_Task12_Images.zip -d imgs/test # For annotations unzip -q Challenge1_Training_Task1_GT.zip -d annotations/training unzip -q Challenge1_Test_Task1_GT.zip -d annotations/test rm Challenge1_Training_Task12_Images.zip && rm Challenge1_Test_Task12_Images.zip && rm Challenge1_Training_Task1_GT.zip && rm Challenge1_Test_Task1_GT.zip
Step 2: Generate
instances_training.json
andinstances_test.json
with the following command:python tools/data/textdet/ic11_converter.py PATH/TO/icdar2011 --nproc 4
After running the above codes, the directory structure should be as follows:
│── icdar2011 │ ├── imgs │ ├── instances_test.json │ └── instances_training.json
ICDAR 2013 (Focused Scene Text)¶
Step1: Download
Challenge2_Training_Task12_Images.zip
,Challenge2_Test_Task12_Images.zip
,Challenge2_Training_Task1_GT.zip
, andChallenge2_Test_Task1_GT.zip
from homepageTask 2.1: Text Localization (2013 edition)
.mkdir icdar2013 && cd icdar2013 mkdir imgs && mkdir annotations # Download ICDAR 2013 wget https://rrc.cvc.uab.es/downloads/Challenge2_Training_Task12_Images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge2_Test_Task12_Images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge2_Training_Task1_GT.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge2_Test_Task1_GT.zip --no-check-certificate # For images unzip -q Challenge2_Training_Task12_Images.zip -d imgs/training unzip -q Challenge2_Test_Task12_Images.zip -d imgs/test # For annotations unzip -q Challenge2_Training_Task1_GT.zip -d annotations/training unzip -q Challenge2_Test_Task1_GT.zip -d annotations/test rm Challenge2_Training_Task12_Images.zip && rm Challenge2_Test_Task12_Images.zip && rm Challenge2_Training_Task1_GT.zip && rm Challenge2_Test_Task1_GT.zip
Step 2: Generate
instances_training.json
andinstances_test.json
with the following command:python tools/data/textdet/ic13_converter.py PATH/TO/icdar2013 --nproc 4
After running the above codes, the directory structure should be as follows:
│── icdar2013 │ ├── imgs │ ├── instances_test.json │ └── instances_training.json
ICDAR 2015¶
Step0: Read Important Note
Step1: Download
ch4_training_images.zip
,ch4_test_images.zip
,ch4_training_localization_transcription_gt.zip
,Challenge4_Test_Task1_GT.zip
from homepageStep2:
mkdir icdar2015 && cd icdar2015 mkdir imgs && mkdir annotations # For images, mv ch4_training_images imgs/training mv ch4_test_images imgs/test # For annotations, mv ch4_training_localization_transcription_gt annotations/training mv Challenge4_Test_Task1_GT annotations/test
Step3: Download instances_training.json and instances_test.json and move them to
icdar2015
Or, generate
instances_training.json
andinstances_test.json
with the following command:python tools/data/textdet/icdar_converter.py /path/to/icdar2015 -o /path/to/icdar2015 -d icdar2015 --split-list training test
The resulting directory structure looks like the following:
├── icdar2015 │ ├── imgs │ ├── annotations │ ├── instances_test.json │ └── instances_training.json
ICDAR 2017¶
Follow similar steps as ICDAR 2015.
The resulting directory structure looks like the following:
├── icdar2017 │ ├── imgs │ ├── annotations │ ├── instances_training.json │ └── instances_val.json
SynthText¶
Step1: Download SynthText.zip from [homepage](https://www.robots.ox.ac.uk/~vgg/data/scenetext/ and extract its content to
synthtext/img
.Step2: Download data.mdb and lock.mdb to
synthtext/instances_training.lmdb/
.The resulting directory structure looks like the following:
├── synthtext │ ├── imgs │ └── instances_training.lmdb │ ├── data.mdb │ └── lock.mdb
TextOCR¶
Step1: Download train_val_images.zip, TextOCR_0.1_train.json and TextOCR_0.1_val.json to
textocr/
.mkdir textocr && cd textocr # Download TextOCR dataset wget https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip wget https://dl.fbaipublicfiles.com/textvqa/data/textocr/TextOCR_0.1_train.json wget https://dl.fbaipublicfiles.com/textvqa/data/textocr/TextOCR_0.1_val.json # For images unzip -q train_val_images.zip mv train_images train
Step2: Generate
instances_training.json
andinstances_val.json
with the following command:python tools/data/textdet/textocr_converter.py /path/to/textocr
The resulting directory structure looks like the following:
├── textocr │ ├── train │ ├── instances_training.json │ └── instances_val.json
Totaltext¶
Step0: Read Important Note
Step1: Download
totaltext.zip
from github dataset andgroundtruth_text.zip
orTT_new_train_GT.zip
(if you prefer to use the latest version of training annotations) from github Groundtruth (Our totaltext_converter.py supports groundtruth with both .mat and .txt format).mkdir totaltext && cd totaltext mkdir imgs && mkdir annotations # For images # in ./totaltext unzip totaltext.zip mv Images/Train imgs/training mv Images/Test imgs/test # For legacy training and test annotations unzip groundtruth_text.zip mv Groundtruth/Polygon/Train annotations/training mv Groundtruth/Polygon/Test annotations/test # Using the latest training annotations # WARNING: Delete legacy train annotations before running the following command. unzip TT_new_train_GT.zip mv Train annotations/training
Step2: Generate
instances_training.json
andinstances_test.json
with the following command:python tools/data/textdet/totaltext_converter.py /path/to/totaltext
The resulting directory structure looks like the following:
├── totaltext │ ├── imgs │ ├── annotations │ ├── instances_test.json │ └── instances_training.json
CurvedSynText150k¶
Step1: Download syntext1.zip and syntext2.zip to
CurvedSynText150k/
.Step2:
unzip -q syntext1.zip mv train.json train1.json unzip images.zip rm images.zip unzip -q syntext2.zip mv train.json train2.json unzip images.zip rm images.zip
Step3: Download instances_training.json to
CurvedSynText150k/
Or, generate
instances_training.json
with following command:python tools/data/common/curvedsyntext_converter.py PATH/TO/CurvedSynText150k --nproc 4
The resulting directory structure looks like the following:
├── CurvedSynText150k │ ├── syntext_word_eng │ ├── emcs_imgs │ └── instances_training.json
FUNSD¶
Step1: Download dataset.zip to
funsd/
.mkdir funsd && cd funsd # Download FUNSD dataset wget https://guillaumejaume.github.io/FUNSD/dataset.zip unzip -q dataset.zip # For images mv dataset/training_data/images imgs && mv dataset/testing_data/images/* imgs/ # For annotations mkdir annotations mv dataset/training_data/annotations annotations/training && mv dataset/testing_data/annotations annotations/test rm dataset.zip && rm -rf dataset
Step2: Generate
instances_training.json
andinstances_test.json
with following command:python tools/data/textdet/funsd_converter.py PATH/TO/funsd --nproc 4
The resulting directory structure looks like the following:
│── funsd │ ├── annotations │ ├── imgs │ ├── instances_test.json │ └── instances_training.json
DeText¶
Step1: Download
ch9_training_images.zip
,ch9_training_localization_transcription_gt.zip
,ch9_validation_images.zip
, andch9_validation_localization_transcription_gt.zip
from Task 3: End to End on the homepage.mkdir detext && cd detext mkdir imgs && mkdir annotations && mkdir imgs/training && mkdir imgs/val && mkdir annotations/training && mkdir annotations/val # Download DeText wget https://rrc.cvc.uab.es/downloads/ch9_training_images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/ch9_training_localization_transcription_gt.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/ch9_validation_images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/ch9_validation_localization_transcription_gt.zip --no-check-certificate # Extract images and annotations unzip -q ch9_training_images.zip -d imgs/training && unzip -q ch9_training_localization_transcription_gt.zip -d annotations/training && unzip -q ch9_validation_images.zip -d imgs/val && unzip -q ch9_validation_localization_transcription_gt.zip -d annotations/val # Remove zips rm ch9_training_images.zip && rm ch9_training_localization_transcription_gt.zip && rm ch9_validation_images.zip && rm ch9_validation_localization_transcription_gt.zip
Step2: Generate
instances_training.json
andinstances_val.json
with following command:python tools/data/textdet/detext_converter.py PATH/TO/detext --nproc 4
After running the above codes, the directory structure should be as follows:
│── detext │ ├── annotations │ ├── imgs │ ├── instances_test.json │ └── instances_training.json
NAF¶
Step1: Download labeled_images.tar.gz to
naf/
.mkdir naf && cd naf # Download NAF dataset wget https://github.com/herobd/NAF_dataset/releases/download/v1.0/labeled_images.tar.gz tar -zxf labeled_images.tar.gz # For images mkdir annotations && mv labeled_images imgs # For annotations git clone https://github.com/herobd/NAF_dataset.git mv NAF_dataset/train_valid_test_split.json annotations/ && mv NAF_dataset/groups annotations/ rm -rf NAF_dataset && rm labeled_images.tar.gz
Step2: Generate
instances_training.json
,instances_val.json
, andinstances_test.json
with following command:python tools/data/textdet/naf_converter.py PATH/TO/naf --nproc 4
After running the above codes, the directory structure should be as follows:
│── naf │ ├── annotations │ ├── imgs │ ├── instances_test.json │ ├── instances_val.json │ └── instances_training.json
SROIE¶
Step1: Download
0325updated.task1train(626p).zip
,task1&2_test(361p).zip
, andtext.task1&2-test(361p).zip
from homepage tosroie/
Step2:
mkdir sroie && cd sroie mkdir imgs && mkdir annotations && mkdir imgs/training # Warnninig: The zip files downloaded from Google Drive and BaiduYun Cloud may # be different, the user should revise the following commands to the correct # file name if encounter with errors while extracting and move the files. unzip -q 0325updated.task1train\(626p\).zip && unzip -q task1\&2_test\(361p\).zip && unzip -q text.task1\&2-test(361p\).zip # For images mv 0325updated.task1train\(626p\)/*.jpg imgs/training && mv fulltext_test\(361p\) imgs/test # For annotations mv 0325updated.task1train\(626p\) annotations/training && mv text.task1\&2-testги361p\)/ annotations/test rm 0325updated.task1train\(626p\).zip && rm task1\&2_test\(361p\).zip && rm text.task1\&2-test(361p\).zip
Step3: Generate
instances_training.json
andinstances_test.json
with the following command:python tools/data/textdet/sroie_converter.py PATH/TO/sroie --nproc 4
After running the above codes, the directory structure should be as follows:
├── sroie │ ├── annotations │ ├── imgs │ ├── instances_test.json │ └── instances_training.json
Lecture Video DB¶
Step1: Download IIIT-CVid.zip to
lv/
.mkdir lv && cd lv # Download LV dataset wget http://cdn.iiit.ac.in/cdn/preon.iiit.ac.in/~kartik/IIIT-CVid.zip unzip -q IIIT-CVid.zip mv IIIT-CVid/Frames imgs rm IIIT-CVid.zip
Step2: Generate
instances_training.json
,instances_val.json
, andinstances_test.json
with following command:python tools/data/textdet/lv_converter.py PATH/TO/lv --nproc 4
The resulting directory structure looks like the following:
│── lv │ ├── imgs │ ├── instances_test.json │ ├── instances_training.json │ └── instances_val.json
LSVT¶
Step1: Download train_full_images_0.tar.gz, train_full_images_1.tar.gz, and train_full_labels.json to
lsvt/
.mkdir lsvt && cd lsvt # Download LSVT dataset wget https://dataset-bj.cdn.bcebos.com/lsvt/train_full_images_0.tar.gz wget https://dataset-bj.cdn.bcebos.com/lsvt/train_full_images_1.tar.gz wget https://dataset-bj.cdn.bcebos.com/lsvt/train_full_labels.json mkdir annotations tar -xf train_full_images_0.tar.gz && tar -xf train_full_images_1.tar.gz mv train_full_labels.json annotations/ && mv train_full_images_1/*.jpg train_full_images_0/ mv train_full_images_0 imgs rm train_full_images_0.tar.gz && rm train_full_images_1.tar.gz && rm -rf train_full_images_1
Step2: Generate
instances_training.json
andinstances_val.json
(optional) with the following command:# Annotations of LSVT test split is not publicly available, split a validation # set by adding --val-ratio 0.2 python tools/data/textdet/lsvt_converter.py PATH/TO/lsvt
After running the above codes, the directory structure should be as follows:
|── lsvt │ ├── imgs │ ├── instances_training.json │ └── instances_val.json (optional)
IMGUR¶
Step1: Run
download_imgur5k.py
to download images. You can merge PR#5 in your local repository to enable a much faster parallel execution of image download.mkdir imgur && cd imgur git clone https://github.com/facebookresearch/IMGUR5K-Handwriting-Dataset.git # Download images from imgur.com. This may take SEVERAL HOURS! python ./IMGUR5K-Handwriting-Dataset/download_imgur5k.py --dataset_info_dir ./IMGUR5K-Handwriting-Dataset/dataset_info/ --output_dir ./imgs # For annotations mkdir annotations mv ./IMGUR5K-Handwriting-Dataset/dataset_info/*.json annotations rm -rf IMGUR5K-Handwriting-Dataset
Step2: Generate
instances_train.json
,instance_val.json
andinstances_test.json
with the following command:python tools/data/textdet/imgur_converter.py PATH/TO/imgur
After running the above codes, the directory structure should be as follows:
│── imgur │ ├── annotations │ ├── imgs │ ├── instances_test.json │ ├── instances_training.json │ └── instances_val.json
KAIST¶
Step1: Complete download KAIST_all.zip to
kaist/
.mkdir kaist && cd kaist mkdir imgs && mkdir annotations # Download KAIST dataset wget http://www.iapr-tc11.org/dataset/KAIST_SceneText/KAIST_all.zip unzip -q KAIST_all.zip rm KAIST_all.zip
Step2: Extract zips:
python tools/data/common/extract_kaist.py PATH/TO/kaist
Step3: Generate
instances_training.json
andinstances_val.json
(optional) with following command:# Since KAIST does not provide an official split, you can split the dataset by adding --val-ratio 0.2 python tools/data/textdet/kaist_converter.py PATH/TO/kaist --nproc 4
After running the above codes, the directory structure should be as follows:
│── kaist │ ├── annotations │ ├── imgs │ ├── instances_training.json │ └── instances_val.json (optional)
MTWI¶
Step1: Download
mtwi_2018_train.zip
from homepage.mkdir mtwi && cd mtwi unzip -q mtwi_2018_train.zip mv image_train imgs && mv txt_train annotations rm mtwi_2018_train.zip
Step2: Generate
instances_training.json
andinstance_val.json
(optional) with the following command:# Annotations of MTWI test split is not publicly available, split a validation # set by adding --val-ratio 0.2 python tools/data/textdet/mtwi_converter.py PATH/TO/mtwi --nproc 4
After running the above codes, the directory structure should be as follows:
│── mtwi │ ├── annotations │ ├── imgs │ ├── instances_training.json │ └── instances_val.json (optional)
COCO Text v2¶
Step1: Download image train2014.zip and annotation cocotext.v2.zip to
coco_textv2/
.mkdir coco_textv2 && cd coco_textv2 mkdir annotations # Download COCO Text v2 dataset wget http://images.cocodataset.org/zips/train2014.zip wget https://github.com/bgshih/cocotext/releases/download/dl/cocotext.v2.zip unzip -q train2014.zip && unzip -q cocotext.v2.zip mv train2014 imgs && mv cocotext.v2.json annotations/ rm train2014.zip && rm -rf cocotext.v2.zip
Step2: Generate
instances_training.json
andinstances_val.json
with the following command:python tools/data/textdet/cocotext_converter.py PATH/TO/coco_textv2
After running the above codes, the directory structure should be as follows:
│── coco_textv2 │ ├── annotations │ ├── imgs │ ├── instances_training.json │ └── instances_val.json
ReCTS¶
Step1: Download ReCTS.zip to
rects/
from the homepage.mkdir rects && cd rects # Download ReCTS dataset # You can also find Google Drive link on the dataset homepage wget https://datasets.cvc.uab.es/rrc/ReCTS.zip --no-check-certificate unzip -q ReCTS.zip mv img imgs && mv gt_unicode annotations rm ReCTS.zip && rm -rf gt
Step2: Generate
instances_training.json
andinstances_val.json
(optional) with following command:# Annotations of ReCTS test split is not publicly available, split a validation # set by adding --val-ratio 0.2 python tools/data/textdet/rects_converter.py PATH/TO/rects --nproc 4 --val-ratio 0.2
After running the above codes, the directory structure should be as follows:
│── rects │ ├── annotations │ ├── imgs │ ├── instances_val.json (optional) │ └── instances_training.json
ILST¶
Step1: Download
IIIT-ILST
from onedriveStep2: Run the following commands
unzip -q IIIT-ILST.zip && rm IIIT-ILST.zip cd IIIT-ILST # rename files cd Devanagari && for i in `ls`; do mv -f $i `echo "devanagari_"$i`; done && cd .. cd Malayalam && for i in `ls`; do mv -f $i `echo "malayalam_"$i`; done && cd .. cd Telugu && for i in `ls`; do mv -f $i `echo "telugu_"$i`; done && cd .. # transfer image path mkdir imgs && mkdir annotations mv Malayalam/{*jpg,*jpeg} imgs/ && mv Malayalam/*xml annotations/ mv Devanagari/*jpg imgs/ && mv Devanagari/*xml annotations/ mv Telugu/*jpeg imgs/ && mv Telugu/*xml annotations/ # remove unnecessary files rm -rf Devanagari && rm -rf Malayalam && rm -rf Telugu && rm -rf README.txt
Step3: Generate
instances_training.json
andinstances_val.json
(optional). Since the original dataset doesn’t have a validation set, you may specify--val-ratio
to split the dataset. E.g., if val-ratio is 0.2, then 20% of the data are left out as the validation set in this example.python tools/data/textdet/ilst_converter.py PATH/TO/IIIT-ILST --nproc 4
After running the above codes, the directory structure should be as follows:
│── IIIT-ILST │ ├── annotations │ ├── imgs │ ├── instances_val.json (optional) │ └── instances_training.json
VinText¶
Step1: Download vintext.zip to
vintext
mkdir vintext && cd vintext # Download dataset from google drive wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1UUQhNvzgpZy7zXBFQp0Qox-BBjunZ0ml' -O- │ sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1UUQhNvzgpZy7zXBFQp0Qox-BBjunZ0ml" -O vintext.zip && rm -rf /tmp/cookies.txt # Extract images and annotations unzip -q vintext.zip && rm vintext.zip mv vietnamese/labels ./ && mv vietnamese/test_image ./ && mv vietnamese/train_images ./ && mv vietnamese/unseen_test_images ./ rm -rf vietnamese # Rename files mv labels annotations && mv test_image test && mv train_images training && mv unseen_test_images unseen_test mkdir imgs mv training imgs/ && mv test imgs/ && mv unseen_test imgs/
Step2: Generate
instances_training.json
,instances_test.json
andinstances_unseen_test.json
python tools/data/textdet/vintext_converter.py PATH/TO/vintext --nproc 4
After running the above codes, the directory structure should be as follows:
│── vintext │ ├── annotations │ ├── imgs │ ├── instances_test.json │ ├── instances_unseen_test.json │ └── instances_training.json
BID¶
Step1: Download BID Dataset.zip
Step2: Run the following commands to preprocess the dataset
# Rename mv BID\ Dataset.zip BID_Dataset.zip # Unzip and Rename unzip -q BID_Dataset.zip && rm BID_Dataset.zip mv BID\ Dataset BID # The BID dataset has a problem of permission, and you may # add permission for this file chmod -R 777 BID cd BID mkdir imgs && mkdir annotations # For images and annotations mv CNH_Aberta/*in.jpg imgs && mv CNH_Aberta/*txt annotations && rm -rf CNH_Aberta mv CNH_Frente/*in.jpg imgs && mv CNH_Frente/*txt annotations && rm -rf CNH_Frente mv CNH_Verso/*in.jpg imgs && mv CNH_Verso/*txt annotations && rm -rf CNH_Verso mv CPF_Frente/*in.jpg imgs && mv CPF_Frente/*txt annotations && rm -rf CPF_Frente mv CPF_Verso/*in.jpg imgs && mv CPF_Verso/*txt annotations && rm -rf CPF_Verso mv RG_Aberto/*in.jpg imgs && mv RG_Aberto/*txt annotations && rm -rf RG_Aberto mv RG_Frente/*in.jpg imgs && mv RG_Frente/*txt annotations && rm -rf RG_Frente mv RG_Verso/*in.jpg imgs && mv RG_Verso/*txt annotations && rm -rf RG_Verso # Remove unnecessary files rm -rf desktop.ini
Step3: - Step3: Generate
instances_training.json
andinstances_val.json
(optional). Since the original dataset doesn’t have a validation set, you may specify--val-ratio
to split the dataset. E.g., if val-ratio is 0.2, then 20% of the data are left out as the validation set in this example.python tools/data/textdet/bid_converter.py PATH/TO/BID --nproc 4
After running the above codes, the directory structure should be as follows:
│── BID │ ├── annotations │ ├── imgs │ ├── instances_training.json │ └── instances_val.json (optional)
RCTW¶
Step1: Download
train_images.zip.001
,train_images.zip.002
, andtrain_gts.zip
from the homepage, extract the zips torctw/imgs
andrctw/annotations
, respectively.Step2: Generate
instances_training.json
andinstances_val.json
(optional). Since the test annotations are not publicly available, you may specify--val-ratio
to split the dataset. E.g., if val-ratio is 0.2, then 20% of the data are left out as the validation set in this example.# Annotations of RCTW test split is not publicly available, split a validation set by adding --val-ratio 0.2 python tools/data/textdet/rctw_converter.py PATH/TO/rctw --nproc 4
After running the above codes, the directory structure should be as follows:
│── rctw │ ├── annotations │ ├── imgs │ ├── instances_training.json │ └── instances_val.json (optional)
HierText¶
Step1 (optional): Install AWS CLI.
Step2: Clone HierText repo to get annotations
mkdir HierText git clone https://github.com/google-research-datasets/hiertext.git
Step3: Download
train.tgz
,validation.tgz
from awsaws s3 --no-sign-request cp s3://open-images-dataset/ocr/train.tgz . aws s3 --no-sign-request cp s3://open-images-dataset/ocr/validation.tgz .
Step4: Process raw data
# process annotations mv hiertext/gt ./ rm -rf hiertext mv gt annotations gzip -d annotations/train.jsonl.gz gzip -d annotations/validation.jsonl.gz # process images mkdir imgs mv train.tgz imgs/ mv validation.tgz imgs/ tar -xzvf imgs/train.tgz tar -xzvf imgs/validation.tgz
Step5: Generate
instances_training.json
andinstance_val.json
. HierText includes different levels of annotation, from paragraph, line, to word. Check the original paper for details. E.g. set--level paragraph
to get paragraph-level annotation. Set--level line
to get line-level annotation. set--level word
to get word-level annotation.# Collect word annotation from HierText --level word python tools/data/textdet/hiertext_converter.py PATH/TO/HierText --level word --nproc 4
After running the above codes, the directory structure should be as follows:
│── HierText │ ├── annotations │ ├── imgs │ ├── instances_training.json │ └── instances_val.json
ArT¶
Step1: Download
train_images.tar.gz
, andtrain_labels.json
from the homepage toart/
mkdir art && cd art mkdir annotations # Download ArT dataset wget https://dataset-bj.cdn.bcebos.com/art/train_images.tar.gz --no-check-certificate wget https://dataset-bj.cdn.bcebos.com/art/train_labels.json --no-check-certificate # Extract tar -xf train_images.tar.gz mv train_images imgs mv train_labels.json annotations/ # Remove unnecessary files rm train_images.tar.gz
Step2: Generate
instances_training.json
andinstances_val.json
(optional). Since the test annotations are not publicly available, you may specify--val-ratio
to split the dataset. E.g., if val-ratio is 0.2, then 20% of the data are left out as the validation set in this example.# Annotations of ArT test split is not publicly available, split a validation set by adding --val-ratio 0.2 python tools/data/textdet/art_converter.py PATH/TO/art --nproc 4
After running the above codes, the directory structure should be as follows:
│── art │ ├── annotations │ ├── imgs │ ├── instances_training.json │ └── instances_val.json (optional)
Text Recognition¶
Overview¶
Dataset | images | annotation file | annotation file |
---|---|---|---|
training | test | ||
coco_text | homepage | train_label.txt | - |
ICDAR2011 | homepage | - | - |
ICDAR2013 | homepage | - | - |
icdar_2015 | homepage | train_label.txt | test_label.txt |
IIIT5K | homepage | train_label.txt | test_label.txt |
ct80 | homepage | - | test_label.txt |
svt | homepage | - | test_label.txt |
svtp | unofficial homepage[1] | - | test_label.txt |
MJSynth (Syn90k) | homepage | shuffle_labels.txt | label.txt | - |
SynthText (Synth800k) | homepage | alphanumeric_labels.txt |shuffle_labels.txt | instances_train.txt | label.txt | - |
SynthAdd | SynthText_Add.zip (code:627x) | label.txt | - |
TextOCR | homepage | - | - |
Totaltext | homepage | - | - |
OpenVINO | Open Images | annotations | annotations |
FUNSD | homepage | - | - |
DeText | homepage | - | - |
NAF | homepage | - | - |
SROIE | homepage | - | - |
Lecture Video DB | homepage | - | - |
LSVT | homepage | - | - |
IMGUR | homepage | - | - |
KAIST | homepage | - | - |
MTWI | homepage | - | - |
COCO Text v2 | homepage | - | - |
ReCTS | homepage | - | - |
IIIT-ILST | homepage | - | - |
VinText | homepage | - | - |
BID | homepage | - | - |
RCTW | homepage | - | - |
HierText | homepage | - | - |
ArT | homepage | - | - |
(*) Since the official homepage is unavailable now, we provide an alternative for quick reference. However, we do not guarantee the correctness of the dataset.
Install AWS CLI (optional)¶
Since there are some datasets that require the AWS CLI to be installed in advance, we provide a quick installation guide here:
curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip" unzip awscliv2.zip sudo ./aws/install ./aws/install -i /usr/local/aws-cli -b /usr/local/bin !aws configure # this command will require you to input keys, you can skip them except # for the Default region name # AWS Access Key ID [None]: # AWS Secret Access Key [None]: # Default region name [None]: us-east-1 # Default output format [None]
ICDAR 2011 (Born-Digital Images)¶
Step1: Download
Challenge1_Training_Task3_Images_GT.zip
,Challenge1_Test_Task3_Images.zip
, andChallenge1_Test_Task3_GT.txt
from homepageTask 1.3: Word Recognition (2013 edition)
.mkdir icdar2011 && cd icdar2011 mkdir annotations # Download ICDAR 2011 wget https://rrc.cvc.uab.es/downloads/Challenge1_Training_Task3_Images_GT.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge1_Test_Task3_Images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge1_Test_Task3_GT.txt --no-check-certificate # For images mkdir crops unzip -q Challenge1_Training_Task3_Images_GT.zip -d crops/train unzip -q Challenge1_Test_Task3_Images.zip -d crops/test # For annotations mv Challenge1_Test_Task3_GT.txt annotations && mv train/gt.txt annotations/Challenge1_Train_Task3_GT.txt
Step2: Convert original annotations to
Train_label.jsonl
andTest_label.jsonl
with the following command:python tools/data/textrecog/ic11_converter.py PATH/TO/icdar2011
After running the above codes, the directory structure should be as follows:
├── icdar2011 │ ├── crops │ ├── train_label.jsonl │ └── test_label.jsonl
ICDAR 2013 (Focused Scene Text)¶
Step1: Download
Challenge2_Training_Task3_Images_GT.zip
,Challenge2_Test_Task3_Images.zip
, andChallenge2_Test_Task3_GT.txt
from homepageTask 2.3: Word Recognition (2013 edition)
.mkdir icdar2013 && cd icdar2013 mkdir annotations # Download ICDAR 2013 wget https://rrc.cvc.uab.es/downloads/Challenge2_Training_Task3_Images_GT.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge2_Test_Task3_Images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/Challenge2_Test_Task3_GT.txt --no-check-certificate # For images mkdir crops unzip -q Challenge2_Training_Task3_Images_GT.zip -d crops/train unzip -q Challenge2_Test_Task3_Images.zip -d crops/test # For annotations mv Challenge2_Test_Task3_GT.txt annotations && mv crops/train/gt.txt annotations/Challenge2_Train_Task3_GT.txt rm Challenge2_Training_Task3_Images_GT.zip && rm Challenge2_Test_Task3_Images.zip
Step 2: Generate
Train_label.jsonl
andTest_label.jsonl
with the following command:python tools/data/textrecog/ic13_converter.py PATH/TO/icdar2013
After running the above codes, the directory structure should be as follows:
├── icdar2013 │ ├── crops │ ├── train_label.jsonl │ └── test_label.jsonl
ICDAR 2013 [Deprecated]¶
Step1: Download
Challenge2_Test_Task3_Images.zip
andChallenge2_Training_Task3_Images_GT.zip
from homepageStep2: Download test_label_1015.txt and train_label.txt
After running the above codes, the directory structure should be as follows:
├── icdar_2013 │ ├── train_label.txt │ ├── test_label_1015.txt │ ├── test_label_1095.txt │ ├── Challenge2_Training_Task3_Images_GT │ └── Challenge2_Test_Task3_Images
ICDAR 2015¶
Step1: Download
ch4_training_word_images_gt.zip
andch4_test_word_images_gt.zip
from homepageStep2: Download train_label.txt and test_label.txt
After running the above codes, the directory structure should be as follows:
├── icdar_2015 │ ├── train_label.txt │ ├── test_label.txt │ ├── ch4_training_word_images_gt │ └── ch4_test_word_images_gt
IIIT5K¶
Step1: Download
IIIT5K-Word_V3.0.tar.gz
from homepageStep2: Download train_label.txt and test_label.txt
After running the above codes, the directory structure should be as follows:
├── III5K │ ├── train_label.txt │ ├── test_label.txt │ ├── train │ └── test
svt¶
Step1: Download
svt.zip
form homepageStep2: Download test_label.txt
Step3:
python tools/data/textrecog/svt_converter.py <download_svt_dir_path>
After running the above codes, the directory structure should be as follows:
├── svt │ ├── test_label.txt │ └── image
ct80¶
Step1: Download test_label.txt
Step2: Download timage.tar.gz
Step3:
mkdir ct80 && cd ct80 mv /path/to/test_label.txt . mv /path/to/timage.tar.gz . tar -xvf timage.tar.gz # create soft link cd /path/to/mmocr/data/mixture ln -s /path/to/ct80 ct80
After running the above codes, the directory structure should be as follows:
├── ct80 │ ├── test_label.txt │ └── timage
svtp¶
Step1: Download test_label.txt
After running the above codes, the directory structure should be as follows:
├── svtp │ ├── test_label.txt │ └── image
coco_text¶
Step1: Download from homepage
Step2: Download train_label.txt
After running the above codes, the directory structure should be as follows:
├── coco_text │ ├── train_label.txt │ └── train_words
MJSynth (Syn90k)¶
Step1: Download
mjsynth.tar.gz
from homepageStep2: Download label.txt (8,919,273 annotations) and shuffle_labels.txt (2,400,000 randomly sampled annotations).
Note
Please make sure you’re using the right annotation to train the model by checking its dataset specs in Model Zoo.
Step3:
mkdir Syn90k && cd Syn90k mv /path/to/mjsynth.tar.gz . tar -xzf mjsynth.tar.gz mv /path/to/shuffle_labels.txt . mv /path/to/label.txt . # create soft link cd /path/to/mmocr/data/mixture ln -s /path/to/Syn90k Syn90k # Convert 'txt' format annos to 'lmdb' (optional) cd /path/to/mmocr python tools/data/utils/lmdb_converter.py data/mixture/Syn90k/label.txt data/mixture/Syn90k/label.lmdb --label-only
After running the above codes, the directory structure should be as follows:
├── Syn90k │ ├── shuffle_labels.txt │ ├── label.txt │ ├── label.lmdb (optional) │ └── mnt
SynthText (Synth800k)¶
Step1: Download
SynthText.zip
from homepageStep2: According to your actual needs, download the most appropriate one from the following options: label.txt (7,266,686 annotations), shuffle_labels.txt (2,400,000 randomly sampled annotations), alphanumeric_labels.txt (7,239,272 annotations with alphanumeric characters only) and instances_train.txt (7,266,686 character-level annotations).
Warning
Please make sure you’re using the right annotation to train the model by checking its dataset specs in Model Zoo.
Step3:
mkdir SynthText && cd SynthText mv /path/to/SynthText.zip . unzip SynthText.zip mv SynthText synthtext mv /path/to/shuffle_labels.txt . mv /path/to/label.txt . mv /path/to/alphanumeric_labels.txt . mv /path/to/instances_train.txt . # create soft link cd /path/to/mmocr/data/mixture ln -s /path/to/SynthText SynthText
Step4: Generate cropped images and labels:
cd /path/to/mmocr python tools/data/textrecog/synthtext_converter.py data/mixture/SynthText/gt.mat data/mixture/SynthText/ data/mixture/SynthText/synthtext/SynthText_patch_horizontal --n_proc 8 # Convert 'txt' format annos to 'lmdb' (optional) cd /path/to/mmocr python tools/data/utils/lmdb_converter.py data/mixture/SynthText/label.txt data/mixture/SynthText/label.lmdb --label-only
After running the above codes, the directory structure should be as follows:
├── SynthText │ ├── alphanumeric_labels.txt │ ├── shuffle_labels.txt │ ├── instances_train.txt │ ├── label.txt │ ├── label.lmdb (optional) │ └── synthtext
SynthAdd¶
Step1: Download
SynthText_Add.zip
from SynthAdd (code:627x))Step2: Download label.txt
Step3:
mkdir SynthAdd && cd SynthAdd mv /path/to/SynthText_Add.zip . unzip SynthText_Add.zip mv /path/to/label.txt . # create soft link cd /path/to/mmocr/data/mixture ln -s /path/to/SynthAdd SynthAdd # Convert 'txt' format annos to 'lmdb' (optional) cd /path/to/mmocr python tools/data/utils/lmdb_converter.py data/mixture/SynthAdd/label.txt data/mixture/SynthAdd/label.lmdb --label-only
After running the above codes, the directory structure should be as follows:
├── SynthAdd │ ├── label.txt │ ├── label.lmdb (optional) │ └── SynthText_Add
Tip
To convert label file from txt
format to lmdb
format,
python tools/data/utils/lmdb_converter.py <txt_label_path> <lmdb_label_path> --label-only
For example,
python tools/data/utils/lmdb_converter.py data/mixture/Syn90k/label.txt data/mixture/Syn90k/label.lmdb --label-only
TextOCR¶
Step1: Download train_val_images.zip, TextOCR_0.1_train.json and TextOCR_0.1_val.json to
textocr/
.mkdir textocr && cd textocr # Download TextOCR dataset wget https://dl.fbaipublicfiles.com/textvqa/images/train_val_images.zip wget https://dl.fbaipublicfiles.com/textvqa/data/textocr/TextOCR_0.1_train.json wget https://dl.fbaipublicfiles.com/textvqa/data/textocr/TextOCR_0.1_val.json # For images unzip -q train_val_images.zip mv train_images train
Step2: Generate
train_label.txt
,val_label.txt
and crop images using 4 processes with the following command:python tools/data/textrecog/textocr_converter.py /path/to/textocr 4
After running the above codes, the directory structure should be as follows:
├── TextOCR │ ├── image │ ├── train_label.txt │ └── val_label.txt
Totaltext¶
Step1: Download
totaltext.zip
from github dataset andgroundtruth_text.zip
orTT_new_train_GT.zip
(if you prefer to use the latest version of training annotations) from github Groundtruth (Our totaltext_converter.py supports groundtruth with both .mat and .txt format).mkdir totaltext && cd totaltext mkdir imgs && mkdir annotations # For images # in ./totaltext unzip totaltext.zip mv Images/Train imgs/training mv Images/Test imgs/test # For legacy training and test annotations unzip groundtruth_text.zip mv Groundtruth/Polygon/Train annotations/training mv Groundtruth/Polygon/Test annotations/test # Using the latest training annotations # WARNING: Delete legacy train annotations before running the following command. unzip TT_new_train_GT.zip mv Train annotations/training
Step2: Generate cropped images,
train_label.txt
andtest_label.txt
with the following command (the cropped images will be saved todata/totaltext/dst_imgs/
):python tools/data/textrecog/totaltext_converter.py /path/to/totaltext
After running the above codes, the directory structure should be as follows:
├── totaltext │ ├── dst_imgs │ ├── train_label.txt │ └── test_label.txt
OpenVINO¶
Step1 (optional): Install AWS CLI.
Step2: Download Open Images subsets
train_1
,train_2
,train_5
,train_f
, andvalidation
toopenvino/
.mkdir openvino && cd openvino # Download Open Images subsets for s in 1 2 5 f; do aws s3 --no-sign-request cp s3://open-images-dataset/tar/train_${s}.tar.gz . done aws s3 --no-sign-request cp s3://open-images-dataset/tar/validation.tar.gz . # Download annotations for s in 1 2 5 f; do wget https://storage.openvinotoolkit.org/repositories/openvino_training_extensions/datasets/open_images_v5_text/text_spotting_openimages_v5_train_${s}.json done wget https://storage.openvinotoolkit.org/repositories/openvino_training_extensions/datasets/open_images_v5_text/text_spotting_openimages_v5_validation.json # Extract images mkdir -p openimages_v5/val for s in 1 2 5 f; do tar zxf train_${s}.tar.gz -C openimages_v5 done tar zxf validation.tar.gz -C openimages_v5/val
Step3: Generate
train_{1,2,5,f}_label.txt
,val_label.txt
and crop images using 4 processes with the following command:python tools/data/textrecog/openvino_converter.py /path/to/openvino 4
After running the above codes, the directory structure should be as follows:
├── OpenVINO │ ├── image_1 │ ├── image_2 │ ├── image_5 │ ├── image_f │ ├── image_val │ ├── train_1_label.txt │ ├── train_2_label.txt │ ├── train_5_label.txt │ ├── train_f_label.txt │ └── val_label.txt
DeText¶
Step1: Download
ch9_training_images.zip
,ch9_training_localization_transcription_gt.zip
,ch9_validation_images.zip
, andch9_validation_localization_transcription_gt.zip
from Task 3: End to End on the homepage.mkdir detext && cd detext mkdir imgs && mkdir annotations && mkdir imgs/training && mkdir imgs/val && mkdir annotations/training && mkdir annotations/val # Download DeText wget https://rrc.cvc.uab.es/downloads/ch9_training_images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/ch9_training_localization_transcription_gt.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/ch9_validation_images.zip --no-check-certificate wget https://rrc.cvc.uab.es/downloads/ch9_validation_localization_transcription_gt.zip --no-check-certificate # Extract images and annotations unzip -q ch9_training_images.zip -d imgs/training && unzip -q ch9_training_localization_transcription_gt.zip -d annotations/training && unzip -q ch9_validation_images.zip -d imgs/val && unzip -q ch9_validation_localization_transcription_gt.zip -d annotations/val # Remove zips rm ch9_training_images.zip && rm ch9_training_localization_transcription_gt.zip && rm ch9_validation_images.zip && rm ch9_validation_localization_transcription_gt.zip
Step2: Generate
instances_training.json
andinstances_val.json
with following command:# Add --preserve-vertical to preserve vertical texts for training, otherwise # vertical images will be filtered and stored in PATH/TO/detext/ignores python tools/data/textrecog/detext_converter.py PATH/TO/detext --nproc 4
After running the above codes, the directory structure should be as follows:
├── detext │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── test_label.jsonl
NAF¶
Step1: Download labeled_images.tar.gz to
naf/
.mkdir naf && cd naf # Download NAF dataset wget https://github.com/herobd/NAF_dataset/releases/download/v1.0/labeled_images.tar.gz tar -zxf labeled_images.tar.gz # For images mkdir annotations && mv labeled_images imgs # For annotations git clone https://github.com/herobd/NAF_dataset.git mv NAF_dataset/train_valid_test_split.json annotations/ && mv NAF_dataset/groups annotations/ rm -rf NAF_dataset && rm labeled_images.tar.gz
Step2: Generate
train_label.txt
,val_label.txt
, andtest_label.txt
with following command:# Add --preserve-vertical to preserve vertical texts for training, otherwise # vertical images will be filtered and stored in PATH/TO/naf/ignores python tools/data/textrecog/naf_converter.py PATH/TO/naf --nproc 4
After running the above codes, the directory structure should be as follows:
├── naf │ ├── crops │ ├── train_label.txt │ ├── val_label.txt │ └── test_label.txt
SROIE¶
Step1: Step1: Download
0325updated.task1train(626p).zip
,task1&2_test(361p).zip
, andtext.task1&2-test(361p).zip
from homepage tosroie/
Step2:
mkdir sroie && cd sroie mkdir imgs && mkdir annotations && mkdir imgs/training # Warnninig: The zip files downloaded from Google Drive and BaiduYun Cloud may # be different, the user should revise the following commands to the correct # file name if encounter with errors while extracting and move the files. unzip -q 0325updated.task1train\(626p\).zip && unzip -q task1\&2_test\(361p\).zip && unzip -q text.task1\&2-test(361p\).zip # For images mv 0325updated.task1train\(626p\)/*.jpg imgs/training && mv fulltext_test\(361p\) imgs/test # For annotations mv 0325updated.task1train\(626p\) annotations/training && mv text.task1\&2-testги361p\)/ annotations/test rm 0325updated.task1train\(626p\).zip && rm task1\&2_test\(361p\).zip && rm text.task1\&2-test(361p\).zip
Step3: Generate
train_label.jsonl
andtest_label.jsonl
and crop images using 4 processes with the following command:python tools/data/textrecog/sroie_converter.py PATH/TO/sroie --nproc 4
After running the above codes, the directory structure should be as follows:
├── sroie │ ├── crops │ ├── train_label.jsonl │ └── test_label.jsonl
Lecture Video DB¶
Note
The LV dataset has already provided cropped images and the corresponding annotations
Step1: Download IIIT-CVid.zip to
lv/
.mkdir lv && cd lv # Download LV dataset wget http://cdn.iiit.ac.in/cdn/preon.iiit.ac.in/~kartik/IIIT-CVid.zip unzip -q IIIT-CVid.zip # For image mv IIIT-CVid/Crops ./ # For annotation mv IIIT-CVid/train.txt train_label.txt && mv IIIT-CVid/val.txt val_label.txt && mv IIIT-CVid/test.txt test_label.txt rm IIIT-CVid.zip
Step2: Generate
train_label.jsonl
,val.jsonl
, andtest.jsonl
with following command:python tools/data/textdreog/lv_converter.py PATH/TO/lv
After running the above codes, the directory structure should be as follows:
├── lv │ ├── Crops │ ├── train_label.jsonl │ └── test_label.jsonl
LSVT¶
Step1: Download train_full_images_0.tar.gz, train_full_images_1.tar.gz, and train_full_labels.json to
lsvt/
.mkdir lsvt && cd lsvt # Download LSVT dataset wget https://dataset-bj.cdn.bcebos.com/lsvt/train_full_images_0.tar.gz wget https://dataset-bj.cdn.bcebos.com/lsvt/train_full_images_1.tar.gz wget https://dataset-bj.cdn.bcebos.com/lsvt/train_full_labels.json mkdir annotations tar -xf train_full_images_0.tar.gz && tar -xf train_full_images_1.tar.gz mv train_full_labels.json annotations/ && mv train_full_images_1/*.jpg train_full_images_0/ mv train_full_images_0 imgs rm train_full_images_0.tar.gz && rm train_full_images_1.tar.gz && rm -rf train_full_images_1
Step2: Generate
train_label.jsonl
andval_label.jsonl
(optional) with the following command:# Annotations of LSVT test split is not publicly available, split a validation # set by adding --val-ratio 0.2 # Add --preserve-vertical to preserve vertical texts for training, otherwise # vertical images will be filtered and stored in PATH/TO/lsvt/ignores python tools/data/textdrecog/lsvt_converter.py PATH/TO/lsvt --nproc 4
After running the above codes, the directory structure should be as follows:
├── lsvt │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── val_label.jsonl (optional)
FUNSD¶
Step1: Download dataset.zip to
funsd/
.mkdir funsd && cd funsd # Download FUNSD dataset wget https://guillaumejaume.github.io/FUNSD/dataset.zip unzip -q dataset.zip # For images mv dataset/training_data/images imgs && mv dataset/testing_data/images/* imgs/ # For annotations mkdir annotations mv dataset/training_data/annotations annotations/training && mv dataset/testing_data/annotations annotations/test rm dataset.zip && rm -rf dataset
Step2: Generate
train_label.txt
andtest_label.txt
and crop images using 4 processes with following command (add--preserve-vertical
if you wish to preserve the images containing vertical texts):python tools/data/textrecog/funsd_converter.py PATH/TO/funsd --nproc 4
After running the above codes, the directory structure should be as follows:
├── funsd │ ├── imgs │ ├── dst_imgs │ ├── annotations │ ├── train_label.txt │ └── test_label.txt
IMGUR¶
Step1: Run
download_imgur5k.py
to download images. You can merge PR#5 in your local repository to enable a much faster parallel execution of image download.mkdir imgur && cd imgur git clone https://github.com/facebookresearch/IMGUR5K-Handwriting-Dataset.git # Download images from imgur.com. This may take SEVERAL HOURS! python ./IMGUR5K-Handwriting-Dataset/download_imgur5k.py --dataset_info_dir ./IMGUR5K-Handwriting-Dataset/dataset_info/ --output_dir ./imgs # For annotations mkdir annotations mv ./IMGUR5K-Handwriting-Dataset/dataset_info/*.json annotations rm -rf IMGUR5K-Handwriting-Dataset
Step2: Generate
train_label.txt
,val_label.txt
andtest_label.txt
and crop images with the following command:python tools/data/textrecog/imgur_converter.py PATH/TO/imgur
After running the above codes, the directory structure should be as follows:
├── imgur │ ├── crops │ ├── train_label.jsonl │ ├── test_label.jsonl │ └── val_label.jsonl
KAIST¶
Step1: Complete download KAIST_all.zip to
kaist/
.mkdir kaist && cd kaist mkdir imgs && mkdir annotations # Download KAIST dataset wget http://www.iapr-tc11.org/dataset/KAIST_SceneText/KAIST_all.zip unzip -q KAIST_all.zip rm KAIST_all.zip
Step2: Extract zips:
python tools/data/common/extract_kaist.py PATH/TO/kaist
Step3: Generate
train_label.jsonl
andval_label.jsonl
(optional) with following command:# Since KAIST does not provide an official split, you can split the dataset by adding --val-ratio 0.2 # Add --preserve-vertical to preserve vertical texts for training, otherwise # vertical images will be filtered and stored in PATH/TO/kaist/ignores python tools/data/textrecog/kaist_converter.py PATH/TO/kaist --nproc 4
After running the above codes, the directory structure should be as follows:
├── kaist │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── val_label.jsonl (optional)
MTWI¶
Step1: Download
mtwi_2018_train.zip
from homepage.mkdir mtwi && cd mtwi unzip -q mtwi_2018_train.zip mv image_train imgs && mv txt_train annotations rm mtwi_2018_train.zip
Step2: Generate
train_label.jsonl
andval_label.jsonl
(optional) with the following command:# Annotations of MTWI test split is not publicly available, split a validation # set by adding --val-ratio 0.2 # Add --preserve-vertical to preserve vertical texts for training, otherwise # vertical images will be filtered and stored in PATH/TO/mtwi/ignores python tools/data/textrecog/mtwi_converter.py PATH/TO/mtwi --nproc 4
After running the above codes, the directory structure should be as follows:
├── mtwi │ ├── crops │ ├── train_label.jsonl │ └── val_label.jsonl (optional)
COCO Text v2¶
Step1: Download image train2014.zip and annotation cocotext.v2.zip to
coco_textv2/
.mkdir coco_textv2 && cd coco_textv2 mkdir annotations # Download COCO Text v2 dataset wget http://images.cocodataset.org/zips/train2014.zip wget https://github.com/bgshih/cocotext/releases/download/dl/cocotext.v2.zip unzip -q train2014.zip && unzip -q cocotext.v2.zip mv train2014 imgs && mv cocotext.v2.json annotations/ rm train2014.zip && rm -rf cocotext.v2.zip
Step2: Generate
train_label.jsonl
andval_label.jsonl
with the following command:# Add --preserve-vertical to preserve vertical texts for training, otherwise # vertical images will be filtered and stored in PATH/TO/mtwi/ignores python tools/data/textrecog/cocotext_converter.py PATH/TO/coco_textv2 --nproc 4
After running the above codes, the directory structure should be as follows:
├── coco_textv2 │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── val_label.jsonl
ReCTS¶
Step1: Download ReCTS.zip to
rects/
from the homepage.mkdir rects && cd rects # Download ReCTS dataset # You can also find Google Drive link on the dataset homepage wget https://datasets.cvc.uab.es/rrc/ReCTS.zip --no-check-certificate unzip -q ReCTS.zip mv img imgs && mv gt_unicode annotations rm ReCTS.zip -f && rm -rf gt
Step2: Generate
train_label.jsonl
andval_label.jsonl
(optional) with the following command:# Annotations of ReCTS test split is not publicly available, split a validation # set by adding --val-ratio 0.2 # Add --preserve-vertical to preserve vertical texts for training, otherwise # vertical images will be filtered and stored in PATH/TO/rects/ignores python tools/data/textrecog/rects_converter.py PATH/TO/rects --nproc 4
After running the above codes, the directory structure should be as follows:
├── rects │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── val_label.jsonl (optional)
ILST¶
Step1: Download
IIIT-ILST.zip
from onedrive linkStep2: Run the following commands
unzip -q IIIT-ILST.zip && rm IIIT-ILST.zip cd IIIT-ILST # rename files cd Devanagari && for i in `ls`; do mv -f $i `echo "devanagari_"$i`; done && cd .. cd Malayalam && for i in `ls`; do mv -f $i `echo "malayalam_"$i`; done && cd .. cd Telugu && for i in `ls`; do mv -f $i `echo "telugu_"$i`; done && cd .. # transfer image path mkdir imgs && mkdir annotations mv Malayalam/{*jpg,*jpeg} imgs/ && mv Malayalam/*xml annotations/ mv Devanagari/*jpg imgs/ && mv Devanagari/*xml annotations/ mv Telugu/*jpeg imgs/ && mv Telugu/*xml annotations/ # remove unnecessary files rm -rf Devanagari && rm -rf Malayalam && rm -rf Telugu && rm -rf README.txt
Step3: Generate
train_label.jsonl
andval_label.jsonl
(optional) and crop images using 4 processes with the following command (add--preserve-vertical
if you wish to preserve the images containing vertical texts). Since the original dataset doesn’t have a validation set, you may specify--val-ratio
to split the dataset. E.g., if val-ratio is 0.2, then 20% of the data are left out as the validation set in this example.python tools/data/textrecog/ilst_converter.py PATH/TO/IIIT-ILST --nproc 4
After running the above codes, the directory structure should be as follows:
├── IIIT-ILST │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── val_label.jsonl (optional)
VinText¶
Step1: Download vintext.zip to
vintext
mkdir vintext && cd vintext # Download dataset from google drive wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1UUQhNvzgpZy7zXBFQp0Qox-BBjunZ0ml' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1UUQhNvzgpZy7zXBFQp0Qox-BBjunZ0ml" -O vintext.zip && rm -rf /tmp/cookies.txt # Extract images and annotations unzip -q vintext.zip && rm vintext.zip mv vietnamese/labels ./ && mv vietnamese/test_image ./ && mv vietnamese/train_images ./ && mv vietnamese/unseen_test_images ./ rm -rf vietnamese # Rename files mv labels annotations && mv test_image test && mv train_images training && mv unseen_test_images unseen_test mkdir imgs mv training imgs/ && mv test imgs/ && mv unseen_test imgs/
Step2: Generate
train_label.jsonl
,test_label.jsonl
,unseen_test_label.jsonl
, and crop images using 4 processes with the following command (add--preserve-vertical
if you wish to preserve the images containing vertical texts).python tools/data/textrecog/vintext_converter.py PATH/TO/vietnamese --nproc 4
After running the above codes, the directory structure should be as follows:
├── vintext │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ ├── test_label.jsonl │ └── unseen_test_label.jsonl
BID¶
Step1: Download BID Dataset.zip
Step2: Run the following commands to preprocess the dataset
# Rename mv BID\ Dataset.zip BID_Dataset.zip # Unzip and Rename unzip -q BID_Dataset.zip && rm BID_Dataset.zip mv BID\ Dataset BID # The BID dataset has a problem of permission, and you may # add permission for this file chmod -R 777 BID cd BID mkdir imgs && mkdir annotations # For images and annotations mv CNH_Aberta/*in.jpg imgs && mv CNH_Aberta/*txt annotations && rm -rf CNH_Aberta mv CNH_Frente/*in.jpg imgs && mv CNH_Frente/*txt annotations && rm -rf CNH_Frente mv CNH_Verso/*in.jpg imgs && mv CNH_Verso/*txt annotations && rm -rf CNH_Verso mv CPF_Frente/*in.jpg imgs && mv CPF_Frente/*txt annotations && rm -rf CPF_Frente mv CPF_Verso/*in.jpg imgs && mv CPF_Verso/*txt annotations && rm -rf CPF_Verso mv RG_Aberto/*in.jpg imgs && mv RG_Aberto/*txt annotations && rm -rf RG_Aberto mv RG_Frente/*in.jpg imgs && mv RG_Frente/*txt annotations && rm -rf RG_Frente mv RG_Verso/*in.jpg imgs && mv RG_Verso/*txt annotations && rm -rf RG_Verso # Remove unnecessary files rm -rf desktop.ini
Step3: Generate
train_label.jsonl
andval_label.jsonl
(optional) and crop images using 4 processes with the following command (add--preserve-vertical
if you wish to preserve the images containing vertical texts). Since the original dataset doesn’t have a validation set, you may specify--val-ratio
to split the dataset. E.g., if test-ratio is 0.2, then 20% of the data are left out as the validation set in this example.python tools/data/textrecog/bid_converter.py dPATH/TO/BID --nproc 4
After running the above codes, the directory structure should be as follows:
├── BID │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── val_label.jsonl (optional)
RCTW¶
Step1: Download
train_images.zip.001
,train_images.zip.002
, andtrain_gts.zip
from the homepage, extract the zips torctw/imgs
andrctw/annotations
, respectively.Step2: Generate
train_label.jsonl
andval_label.jsonl
(optional). Since the original dataset doesn’t have a validation set, you may specify--val-ratio
to split the dataset. E.g., if val-ratio is 0.2, then 20% of the data are left out as the validation set in this example.# Annotations of RCTW test split is not publicly available, split a validation set by adding --val-ratio 0.2 # Add --preserve-vertical to preserve vertical texts for training, otherwise vertical images will be filtered and stored in PATH/TO/rctw/ignores python tools/data/textrecog/rctw_converter.py PATH/TO/rctw --nproc 4
After running the above codes, the directory structure should be as follows:
│── rctw │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── val_label.jsonl (optional)
HierText¶
Step1 (optional): Install AWS CLI.
Step2: Clone HierText repo to get annotations
mkdir HierText git clone https://github.com/google-research-datasets/hiertext.git
Step3: Download
train.tgz
,validation.tgz
from awsaws s3 --no-sign-request cp s3://open-images-dataset/ocr/train.tgz . aws s3 --no-sign-request cp s3://open-images-dataset/ocr/validation.tgz .
Step4: Process raw data
# process annotations mv hiertext/gt ./ rm -rf hiertext mv gt annotations gzip -d annotations/train.jsonl.gz gzip -d annotations/validation.jsonl.gz # process images mkdir imgs mv train.tgz imgs/ mv validation.tgz imgs/ tar -xzvf imgs/train.tgz tar -xzvf imgs/validation.tgz
Step5: Generate
train_label.jsonl
andval_label.jsonl
. HierText includes different levels of annotation, includingparagraph
,line
, andword
. Check the original paper for details. E.g. set--level paragraph
to get paragraph-level annotation. Set--level line
to get line-level annotation. set--level word
to get word-level annotation.# Collect word annotation from HierText --level word # Add --preserve-vertical to preserve vertical texts for training, otherwise vertical images will be filtered and stored in PATH/TO/HierText/ignores python tools/data/textrecog/hiertext_converter.py PATH/TO/HierText --level word --nproc 4
After running the above codes, the directory structure should be as follows:
│── HierText │ ├── crops │ ├── ignores │ ├── train_label.jsonl │ └── val_label.jsonl
ArT¶
Step1: Download
train_images.tar.gz
, andtrain_labels.json
from the homepage toart/
mkdir art && cd art mkdir annotations # Download ArT dataset wget https://dataset-bj.cdn.bcebos.com/art/train_task2_images.tar.gz wget https://dataset-bj.cdn.bcebos.com/art/train_task2_labels.json # Extract tar -xf train_task2_images.tar.gz mv train_task2_images crops mv train_task2_labels.json annotations/ # Remove unnecessary files rm train_images.tar.gz
Step2: Generate
train_label.jsonl
andval_label.jsonl
(optional). Since the test annotations are not publicly available, you may specify--val-ratio
to split the dataset. E.g., if val-ratio is 0.2, then 20% of the data are left out as the validation set in this example.# Annotations of ArT test split is not publicly available, split a validation set by adding --val-ratio 0.2 python tools/data/textrecog/art_converter.py PATH/TO/art
After running the above codes, the directory structure should be as follows:
│── art │ ├── crops │ ├── train_label.jsonl │ └── val_label.jsonl (optional)
Key Information Extraction¶
Overview¶
The structure of the key information extraction dataset directory is organized as follows.
└── wildreceipt
├── class_list.txt
├── dict.txt
├── image_files
├── openset_train.txt
├── openset_test.txt
├── test.txt
└── train.txt
Preparation Steps¶
WildReceipt¶
Just download and extract wildreceipt.tar.
WildReceiptOpenset¶
Step0: have WildReceipt prepared.
Step1: Convert annotation files to OpenSet format:
# You may find more available arguments by running
# python tools/data/kie/closeset_to_openset.py -h
python tools/data/kie/closeset_to_openset.py data/wildreceipt/train.txt data/wildreceipt/openset_train.txt
python tools/data/kie/closeset_to_openset.py data/wildreceipt/test.txt data/wildreceipt/openset_test.txt
Note
You can learn more about the key differences between CloseSet and OpenSet annotations in our tutorial.
Named Entity Recognition¶
Overview¶
The structure of the named entity recognition dataset directory is organized as follows.
└── cluener2020
├── cluener_predict.json
├── dev.json
├── README.md
├── test.json
├── train.json
└── vocab.txt
Preparation Steps¶
CLUENER2020¶
Download and extract cluener_public.zip to
cluener2020/
Download vocab.txt and move
vocab.txt
tocluener2020/
Useful Tools¶
We provide some useful tools under mmocr/tools
directory.
Publish a Model¶
Before you upload a model to AWS, you may want to
(1) convert the model weights to CPU tensors, (2) delete the optimizer states and
(3) compute the hash of the checkpoint file and append the hash id to the filename. These functionalities could be achieved by tools/publish_model.py
.
python tools/publish_model.py ${INPUT_FILENAME} ${OUTPUT_FILENAME}
For example,
python tools/publish_model.py work_dirs/psenet/latest.pth psenet_r50_fpnf_sbn_1x_20190801.pth
The final output filename will be psenet_r50_fpnf_sbn_1x_20190801-{hash id}.pth
.
Convert text recognition dataset to lmdb format¶
Reading images or labels from files can be slow when data are excessive, e.g. on a scale of millions. Besides, in academia, most of the scene text recognition datasets are stored in lmdb format, including images and labels. To get closer to the mainstream practice and enhance the data storage efficiency, MMOCR now provides tools/data/utils/lmdb_converter.py
to convert text recognition datasets to lmdb format.
Arguments | Type | Description |
---|---|---|
label_path |
str | Path to label file. |
output |
str | Output lmdb path. |
--img-root |
str | Input imglist path. |
--label-only |
bool | Only converter label to lmdb |
--label-format |
str | The format of the label file, either txt or jsonl. |
--batch-size |
int | Processing batch size, defaults to 1000 |
--encoding |
str | Bytes coding scheme, defaults to utf8. |
--lmdb-map-size |
int | Maximum size database may grow to , defaults to 1099511627776 bytes (1TB) |
Examples¶
Generate a mixed lmdb file with label.txt and images in imgs/
:
python tools/data/utils/lmdb_converter.py label.txt imgs.lmdb -i imgs
Generate a mixed lmdb file with label.jsonl and images in imgs/
:
python tools/data/utils/lmdb_converter.py label.json imgs.lmdb -i imgs -f jsonl
Generate a label-only lmdb file with label.txt:
python tools/data/utils/lmdb_converter.py label.txt label.lmdb --label-only
Generate a label-only lmdb file with label.jsonl:
python tools/data/utils/lmdb_converter.py label.json label.lmdb --label-only -f jsonl
Convert annotations from Labelme¶
Labelme is a popular graphical image annotation tool. You can convert the labels generated by labelme to the MMOCR data format using tools/data/common/labelme_converter.py
. Both detection and recognition tasks are supported.
# tasks can be "det" or both "det", "recog"
python tools/data/common/labelme_converter.py <json_dir> <image_dir> <out_dir> --tasks <tasks>
For example, converting the labelme format annotation in tests/data/toy_dataset/labelme
to MMOCR detection labels instances_training.txt
and cropping the image patches for recognition task to tests/data/toy_dataset/crops
with the labels train_label.jsonl
:
python tools/data/common/labelme_converter.py tests/data/toy_dataset/labelme tests/data/toy_dataset/imgs tests/data/toy_dataset --tasks det recog
Log Analysis¶
You can use tools/analyze_logs.py
to plot loss/hmean curves given a training log file. Run pip install seaborn
first to install the dependency.
python tools/analyze_logs.py plot_curve [--keys ${KEYS}] [--title ${TITLE}] [--legend ${LEGEND}] [--backend ${BACKEND}] [--style ${STYLE}] [--out ${OUT_FILE}]
Arguments | Type | Description |
---|---|---|
--keys |
str | The metric that you want to plot. Defaults to loss . |
--title |
str | Title of figure. |
--legend |
str | Legend of each plot. |
--backend |
str | Backend of the plot. more info |
--style |
str | Style of the plot. Defaults to dark . more info |
--out |
str | Path of output figure. |
Examples:
Download the following DBNet and CRNN training logs to run demos.
wget https://download.openmmlab.com/mmocr/textdet/dbnet/dbnet_r18_fpnc_sbn_1200e_icdar2015_20210329-ba3ab597.log.json -O DBNet_log.json
wget https://download.openmmlab.com/mmocr/textrecog/crnn/20210326_111035.log.json -O CRNN_log.json
Please specify an output path if you are running the codes on systems without a GUI.
Plot loss metric.
python tools/analyze_logs.py plot_curve DBNet_log.json --keys loss --legend loss
Plot hmean-iou:hmean metric of text detection.
python tools/analyze_logs.py plot_curve DBNet_log.json --keys hmean-iou:hmean --legend hmean-iou:hmean
Plot 0_1-N.E.D metric of text recognition.
python tools/analyze_logs.py plot_curve CRNN_log.json --keys 0_1-N.E.D --legend 0_1-N.E.D
Compute the average training speed.
python tools/analyze_logs.py cal_train_time CRNN_log.json --include-outliers
The output is expected to be like the following.
-----Analyze train time of CRNN_log.json----- slowest epoch 4, average time is 0.3464 fastest epoch 5, average time is 0.2365 time std over epochs is 0.0356 average iter time: 0.2906 s/iter
Changelog¶
0.6.3 (03/11/2022)¶
Highlights¶
This release enhances the inference script and fixes a bug that might cause failure on TorchServe.
Besides, a new backbone, oCLIP-ResNet, and a dataset preparation tool, Dataset Preparer, have been released in MMOCR 1.0.0rc3 (1.x branch). Check out the changelog for more information about the features, and maintenance plan for how we will maintain MMOCR in the future.
New Features & Enhancements¶
Convert numpy.float32 type to python built-in float type by @JunYao1020 in https://github.com/open-mmlab/mmocr/pull/1462
When ‘.’ char not in output string, output is also considered to be a… by @JunYao1020 in https://github.com/open-mmlab/mmocr/pull/1457
Refactor issue template by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/1449
issue template by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/1489
Update maintainers by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1504
Support MMCV < 1.8.0 by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1508
Bug Fixes¶
fix ci by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/1491
[CI] Fix CI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1463
Docs¶
[DOCs] Add MMYOLO in Readme. by @ysh329 in https://github.com/open-mmlab/mmocr/pull/1475
[Docs] Update contributing.md by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1490
New Contributors¶
@ysh329 made their first contribution in https://github.com/open-mmlab/mmocr/pull/1475
Full Changelog: https://github.com/open-mmlab/mmocr/compare/v0.6.2…v0.6.3
0.6.2 (14/10/2022)¶
Highlights¶
It’s now possible to train/test models through Python Interface. For example, you can train a model under mmocr/ directory in this way:
# an example of how to use such modifications is shown as the following:
from mmocr.tools.train import TrainArg, parse_args, run_train_cmd
args = TrainArg(config='/path/to/config.py')
args.add_arg('--work-dir', '/path/to/dir')
args = parse_args(args.arg_list)
run_train_cmd(args)
See PR #1138 for more details.
Besides, release candidates for MMOCR 1.0 with tons of new features are available at 1.x branch now! Check out the changelog for more information about the features, and maintenance plan for how we will maintain MMOCR in the future.
New Features¶
Adding test & train API to be used directly in code by @wybryan in https://github.com/open-mmlab/mmocr/pull/1138
Let ResizeOCR full support mmcv.impad’s pad_val parameters by @hsiehpinghan in https://github.com/open-mmlab/mmocr/pull/1437
Bug Fixes¶
Fix ABINet config by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1256
Fix Recognition Score Normalization Issue by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/1333
Remove max_seq_len inconsistency by @antoniolanza1996 in https://github.com/open-mmlab/mmocr/pull/1433
box points ordering by @yjmm10 in https://github.com/open-mmlab/mmocr/pull/1205
Correct spelling by misspelling ‘preperties’ to ‘properties’ by @JunYao1020 in https://github.com/open-mmlab/mmocr/pull/1446
Docs¶
Demo, experiments and live inference API on Tiyaro by @Venkat2811 in https://github.com/open-mmlab/mmocr/pull/1272
Update 1.x info by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/1369
Add global notes to the docs and the version switcher menu by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1406
Logger Hook Config Updated to Add WandB by @Nourollah in https://github.com/open-mmlab/mmocr/pull/1345
New Contributors¶
@Venkat2811 made their first contribution in https://github.com/open-mmlab/mmocr/pull/1272
@wybryan made their first contribution in https://github.com/open-mmlab/mmocr/pull/1139
@hsiehpinghan made their first contribution in https://github.com/open-mmlab/mmocr/pull/1437
@yjmm10 made their first contribution in https://github.com/open-mmlab/mmocr/pull/1205
@JunYao1020 made their first contribution in https://github.com/open-mmlab/mmocr/pull/1446
@Nourollah made their first contribution in https://github.com/open-mmlab/mmocr/pull/1345
Full Changelog: https://github.com/open-mmlab/mmocr/compare/v0.6.1…v0.6.2
0.6.1 (04/08/2022)¶
Highlights¶
New Features & Enhancements¶
Add ArT by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/1006
add ABINet_Vision api by @Abdelrahman350 in https://github.com/open-mmlab/mmocr/pull/1041
add codespell ignore and use mdformat by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/1022
Add mim to extras_requrie to setup.py, update mminstall… by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1062
Simplify normalized edit distance calculation by @maxbachmann in https://github.com/open-mmlab/mmocr/pull/1060
Test mim in CI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1090
Remove redundant steps by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1091
Update links to SDMGR links by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1252
Bug Fixes¶
Remove unnecessary requirements by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1000
Remove confusing img_scales in pipelines by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1007
inplace operator “+=” will cause RuntimeError when model backward by @garvan2021 in https://github.com/open-mmlab/mmocr/pull/1018
Fix a typo problem in MASTER by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/1031
Fix config name of MASTER in ocr.py by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/1044
Relax OpenCV requirement by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1061
Restrict the minimum version of OpenCV to avoid potential vulnerability by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1065
typo by @tpoisonooo in https://github.com/open-mmlab/mmocr/pull/1024
Fix a typo in setup.py by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1095
fix #1067: add torchserve DockerFile and fix bugs by @Hegelim in https://github.com/open-mmlab/mmocr/pull/1073
Incorrect filename in labelme_converter.py by @xiefeifeihu in https://github.com/open-mmlab/mmocr/pull/1103
Fix dataset configs by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/1106
Fix #1098: normalize text recognition scores by @Hegelim in https://github.com/open-mmlab/mmocr/pull/1119
Update ST_SA_MJ_train.py by @MingyuLau in https://github.com/open-mmlab/mmocr/pull/1117
PSENet metafile by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1121
Flexible ways of getting file name by @balandongiv in https://github.com/open-mmlab/mmocr/pull/1107
Updating edge-embeddings after each GNN layer by @amitbcp in https://github.com/open-mmlab/mmocr/pull/1134
links update by @TekayaNidham in https://github.com/open-mmlab/mmocr/pull/1141
bug fix: access params by cfg.get by @doem97 in https://github.com/open-mmlab/mmocr/pull/1145
Fix a bug in LmdbAnnFileBackend that cause breaking in Synthtext detection training by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/1159
Fix typo of –lmdb-map-size default value by @easilylazy in https://github.com/open-mmlab/mmocr/pull/1147
Fixed docstring syntax error of line 19 & 21 by @APX103 in https://github.com/open-mmlab/mmocr/pull/1157
Update lmdb_converter and ct80 cropped image source in document by @doem97 in https://github.com/open-mmlab/mmocr/pull/1164
MMCV compatibility due to outdated MMDet by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1192
Update maximum version of mmcv by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/1219
Update ABINet links for main by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/1221
Update owners by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1248
Add back some missing fields in configs by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1171
Docs¶
Fix typos by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/1001
Configure Myst-parser to parse anchor tag by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1012
Fix a error in docs/en/tutorials/dataset_types.md by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/1034
Update readme according to the guideline by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1047
Limit markdown version by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1172
Limit extension versions by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/1210
Update installation guide by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1254
Update image link @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/1255
New Contributors¶
@tpoisonooo made their first contribution in https://github.com/open-mmlab/mmocr/pull/1024
@Abdelrahman350 made their first contribution in https://github.com/open-mmlab/mmocr/pull/1041
@Hegelim made their first contribution in https://github.com/open-mmlab/mmocr/pull/1073
@xiefeifeihu made their first contribution in https://github.com/open-mmlab/mmocr/pull/1103
@MingyuLau made their first contribution in https://github.com/open-mmlab/mmocr/pull/1117
@balandongiv made their first contribution in https://github.com/open-mmlab/mmocr/pull/1107
@amitbcp made their first contribution in https://github.com/open-mmlab/mmocr/pull/1134
@TekayaNidham made their first contribution in https://github.com/open-mmlab/mmocr/pull/1141
@easilylazy made their first contribution in https://github.com/open-mmlab/mmocr/pull/1147
@APX103 made their first contribution in https://github.com/open-mmlab/mmocr/pull/1157
Full Changelog: https://github.com/open-mmlab/mmocr/compare/v0.6.0…v0.6.1
0.6.0 (05/05/2022)¶
Highlights¶
A new recognition algorithm MASTER has been added into MMOCR, which was the championship solution for the “ICDAR 2021 Competition on Scientific Table Image Recognition to Latex”! The model pre-trained on SynthText and MJSynth is available for testing! Credit to @JiaquanYe
DBNet++ has been released now! A new Adaptive Scale Fusion module has been equipped for feature enhancement. Benefiting from this, the new model achieved 2% better h-mean score than its predecessor on the ICDAR2015 dataset.
Three more dataset converters are added: LSVT, RCTW and HierText. Check the dataset zoo (Det & Recog ) to explore further information.
To enhance the data storage efficiency, MMOCR now supports loading both images and labels from .lmdb format annotations for the text recognition task. To enable such a feature, the new lmdb_converter.py is ready for use to pack your cropped images and labels into an lmdb file. For a detailed tutorial, please refer to the following sections and the doc.
Testing models on multiple datasets is a widely used evaluation strategy. MMOCR now supports automatically reporting mean scores when there is more than one dataset to evaluate, which enables a more convenient comparison between checkpoints. Doc
Evaluation is more flexible and customizable now. For text detection tasks, you can set the score threshold range where the best results might come out. (Doc) If too many results are flooding your text recognition train log, you can trim it by specifying a subset of metrics in evaluation config. Check out the Evaluation section for details.
MMOCR provides a script to convert the .json labels obtained by the popular annotation toolkit Labelme to MMOCR-supported data format. @Y-M-Y contributed a log analysis tool that helps users gain a better understanding of the entire training process. Read tutorial docs to get started.
Lmdb Dataset¶
Reading images or labels from files can be slow when data are excessive, e.g. on a scale of millions. Besides, in academia, most of the scene text recognition datasets are stored in lmdb format, including images and labels. To get closer to the mainstream practice and enhance the data storage efficiency, MMOCR now officially supports loading images and labels from lmdb datasets via a new pipeline LoadImageFromLMDB. This section is intended to serve as a quick walkthrough for you to master this update and apply it to facilitate your research.
Specifications¶
To better align with the academic community, MMOCR now requires the following specifications for lmdb datasets:
The parameter describing the data volume of the dataset is
num-samples
instead oftotal_number
(deprecated).Images and labels are stored with keys in the form of
image-000000001
andlabel-000000001
, respectively.
Usage¶
Use existing academic lmdb datasets if they meet the specifications; or the tool provided by MMOCR to pack images & annotations into a lmdb dataset.
Previously, MMOCR had a function
txt2lmdb
(deprecated) that only supported converting labels to lmdb format. However, it is quite different from academic lmdb datasets, which usually contain both images and labels. Now MMOCR provides a new utility lmdb_converter to convert recognition datasets with both images and labels to lmdb format.Say that your recognition data in MMOCR’s format are organized as follows. (See an example in ocr_toy_dataset).
# Directory structure ├──img_path | |—— img1.jpg | |—— img2.jpg | |—— ... |——label.txt (or label.jsonl) # Annotation format label.txt: img1.jpg HELLO img2.jpg WORLD ... label.jsonl: {'filename':'img1.jpg', 'text':'HELLO'} {'filename':'img2.jpg', 'text':'WORLD'} ...
Then pack these files up:
python tools/data/utils/lmdb_converter.py {PATH_TO_LABEL} {OUTPUT_PATH} --i {PATH_TO_IMAGES}
Check out tools.md for more details.
The second step is to modify the configuration files. For example, to train CRNN on MJ and ST datasets:
Set parser as
LineJsonParser
andfile_format
as ‘lmdb’ in dataset config# configs/_base_/recog_datasets/ST_MJ_train.py train1 = dict( type='OCRDataset', img_prefix=train_img_prefix1, ann_file=train_ann_file1, loader=dict( type='AnnFileLoader', repeat=1, file_format='lmdb', parser=dict( type='LineJsonParser', keys=['filename', 'text'], )), pipeline=None, test_mode=False)
Use
LoadImageFromLMDB
in pipeline:# configs/_base_/recog_pipelines/crnn_pipeline.py train_pipeline = [ dict(type='LoadImageFromLMDB', color_type='grayscale'), ...
You are good to go! Start training and MMOCR will load data from your lmdb dataset.
New Features & Enhancements¶
Add analyze_logs in tools and its description in docs by @Y-M-Y in https://github.com/open-mmlab/mmocr/pull/899
Add LSVT Data Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/896
Add RCTW dataset converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/914
Support computing mean scores in UniformConcatDataset by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/981
Support loading images and labels from lmdb file by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/982
Add recog2lmdb and new toy dataset files by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/979
Add labelme converter for textdet and textrecog by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/972
Update CircleCI configs by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/918
Update Git Action by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/930
More customizable fields in dataloaders by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/933
Skip CIs when docs are modified by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/941
Rename Github tests, fix ignored paths by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/946
Support latest MMCV by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/959
Support dynamic threshold range in eval_hmean by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/962
Update the version requirement of mmdet in docker by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/966
Replace
opencv-python-headless
withopen-python
by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/970Update Dataset Configs by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/980
Add SynthText dataset config by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/983
Automatically report mean scores when applicable by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/995
Add DBNet++ by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/973
Add MASTER by @JiaquanYe in https://github.com/open-mmlab/mmocr/pull/807
Allow choosing metrics to report in text recognition tasks by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/989
Add HierText converter by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/948
Fix lint_only in CircleCI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/998
Bug Fixes¶
Fix CircleCi Main Branch Accidentally Run PR Stage Test by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/927
Fix a deprecate warning about mmdet.datasets.pipelines.formating by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/944
Fix a Bug in ResNet plugin by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/967
revert a wrong setting in db_r18 cfg by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/978
Fix TotalText Anno version issue by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/945
Update installation step of
albumentations
by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/984Fix ImgAug transform by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/949
Fix GPG key error in CI and docker by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/988
update label.lmdb by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/991
correct meta key by @garvan2021 in https://github.com/open-mmlab/mmocr/pull/926
Use new image by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/976
Fix Data Converter Issues by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/955
Docs¶
Update CONTRIBUTING.md by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/905
Fix the misleading description in test.py by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/908
Update recog.md for lmdb Generation by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/934
Add MMCV by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/954
Add wechat QR code to CN readme by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/960
Update CONTRIBUTING.md by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/947
Use QR codes from MMCV by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/971
Renew dataset_types.md by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/997
New Contributors¶
@Y-M-Y made their first contribution in https://github.com/open-mmlab/mmocr/pull/899
Full Changelog: https://github.com/open-mmlab/mmocr/compare/v0.5.0…v0.6.0
0.5.0 (31/03/2022)¶
Highlights¶
MMOCR now supports SPACE recognition! (What a prominent feature!) Users only need to convert the recognition annotations that contain spaces from a plain
.txt
file to JSON line format.jsonl
, and then revise a few configurations to enable theLineJsonParser
. For more information, please read our step-by-step tutorial.Tesseract is now available in MMOCR! While MMOCR is more flexible to support various downstream tasks, users might sometimes not be satisfied with DL models and would like to turn to effective legacy solutions. Therefore, we offer this option in
mmocr.utils.ocr
by wrapping Tesseract as a detector and/or recognizer. Users can easily create an MMOCR object byMMOCR(det=’Tesseract’, recog=’Tesseract’)
. Credit to @garvan2021We release data converters for 16 widely used OCR datasets, including multiple scenarios such as document, handwritten, and scene text. Now it is more convenient to generate annotation files for these datasets. Check the dataset zoo ( Det & Recog ) to explore further information.
Special thanks to @EighteenSprings @BeyondYourself @yangrisheng, who had actively participated in documentation translation!
Migration Guide - ResNet¶
Some refactoring processes are still going on. For text recognition models, we unified the ResNet-like
architectures which are used as backbones. By introducing stage-wise and block-wise plugins, the refactored ResNet is highly flexible to support existing models, like ResNet31 and ResNet45, and other future designs of ResNet variants.
Plugin¶
Plugin
is a module category inherited from MMCV’s implementation ofPLUGIN_LAYERS
, which can be inserted between each stage of ResNet or into a basicblock. You can find a simple implementation of plugin at mmocr/models/textrecog/plugins/common.py, or click the button below.Plugin Example
@PLUGIN_LAYERS.register_module() class Maxpool2d(nn.Module): """A wrapper around nn.Maxpool2d(). Args: kernel_size (int or tuple(int)): Kernel size for max pooling layer stride (int or tuple(int)): Stride for max pooling layer padding (int or tuple(int)): Padding for pooling layer """ def __init__(self, kernel_size, stride, padding=0, **kwargs): super(Maxpool2d, self).__init__() self.model = nn.MaxPool2d(kernel_size, stride, padding) def forward(self, x): """ Args: x (Tensor): Input feature map Returns: Tensor: The tensor after Maxpooling layer. """ return self.model(x)
Stage-wise Plugins¶
ResNet is composed of stages, and each stage is composed of blocks. E.g., ResNet18 is composed of 4 stages, and each stage is composed of basicblocks. For each stage, we provide two ports to insert stage-wise plugins by giving
plugins
parameters in ResNet.[port1: before stage] ---> [stage] ---> [port2: after stage]
E.g. Using a ResNet with four stages as example. Suppose we want to insert an additional convolution layer before each stage, and an additional convolution layer at stage 1, 2, 4. Then you can define the special ResNet18 like this
resnet18_speical = ResNet( # for simplicity, some required # parameters are omitted plugins=[ dict( cfg=dict( type='ConvModule', kernel_size=3, stride=1, padding=1, norm_cfg=dict(type='BN'), act_cfg=dict(type='ReLU')), stages=(True, True, True, True), position='before_stage') dict( cfg=dict( type='ConvModule', kernel_size=3, stride=1, padding=1, norm_cfg=dict(type='BN'), act_cfg=dict(type='ReLU')), stages=(True, True, False, True), position='after_stage') ])
You can also insert more than one plugin in each port and those plugins will be executed in order. Let’s take ResNet in MASTER as an example:
Multiple Plugins Example
ResNet in Master is based on ResNet31. And after each stage, a module named
GCAModule
will be used. TheGCAModule
is inserted before the stage-wise convolution layer in ResNet31. In conlusion, there will be two plugins atafter_stage
port in the same time.resnet_master = ResNet( # for simplicity, some required # parameters are omitted plugins=[ dict( cfg=dict(type='Maxpool2d', kernel_size=2, stride=(2, 2)), stages=(True, True, False, False), position='before_stage'), dict( cfg=dict(type='Maxpool2d', kernel_size=(2, 1), stride=(2, 1)), stages=(False, False, True, False), position='before_stage'), dict( cfg=dict(type='GCAModule', kernel_size=3, stride=1, padding=1), stages=[True, True, True, True], position='after_stage'), dict( cfg=dict( type='ConvModule', kernel_size=3, stride=1, padding=1, norm_cfg=dict(type='BN'), act_cfg=dict(type='ReLU')), stages=(True, True, True, True), position='after_stage') ])
In each plugin, we will pass two parameters (
in_channels
,out_channels
) to support operations that need the information of current channels.
Block-wise Plugin (Experimental)¶
We also refactored the
BasicBlock
used in ResNet. Now it can be customized with block-wise plugins. Check here for more details.BasicBlock is composed of two convolution layer in the main branch and a shortcut branch. We provide four ports to insert plugins.
[port1: before_conv1] ---> [conv1] ---> [port2: after_conv1] ---> [conv2] ---> [port3: after_conv2] ---> +(shortcut) ---> [port4: after_shortcut]
In each plugin, we will pass a parameter
in_channels
to support operations that need the information of current channels.E.g. Build a ResNet with customized BasicBlock with an additional convolution layer before conv1:
Block-wise Plugin Example
resnet_31 = ResNet( in_channels=3, stem_channels=[64, 128], block_cfgs=dict(type='BasicBlock'), arch_layers=[1, 2, 5, 3], arch_channels=[256, 256, 512, 512], strides=[1, 1, 1, 1], plugins=[ dict( cfg=dict(type='Maxpool2d', kernel_size=2, stride=(2, 2)), stages=(True, True, False, False), position='before_stage'), dict( cfg=dict(type='Maxpool2d', kernel_size=(2, 1), stride=(2, 1)), stages=(False, False, True, False), position='before_stage'), dict( cfg=dict( type='ConvModule', kernel_size=3, stride=1, padding=1, norm_cfg=dict(type='BN'), act_cfg=dict(type='ReLU')), stages=(True, True, True, True), position='after_stage') ])
Full Examples¶
ResNet without plugins
ResNet45 is used in ASTER and ABINet without any plugins.
resnet45_aster = ResNet( in_channels=3, stem_channels=[64, 128], block_cfgs=dict(type='BasicBlock', use_conv1x1='True'), arch_layers=[3, 4, 6, 6, 3], arch_channels=[32, 64, 128, 256, 512], strides=[(2, 2), (2, 2), (2, 1), (2, 1), (2, 1)]) resnet45_abi = ResNet( in_channels=3, stem_channels=32, block_cfgs=dict(type='BasicBlock', use_conv1x1='True'), arch_layers=[3, 4, 6, 6, 3], arch_channels=[32, 64, 128, 256, 512], strides=[2, 1, 2, 1, 1])
ResNet with plugins
ResNet31 is a typical architecture to use stage-wise plugins. Before the first three stages, Maxpooling layer is used. After each stage, a convolution layer with BN and ReLU is used.
resnet_31 = ResNet( in_channels=3, stem_channels=[64, 128], block_cfgs=dict(type='BasicBlock'), arch_layers=[1, 2, 5, 3], arch_channels=[256, 256, 512, 512], strides=[1, 1, 1, 1], plugins=[ dict( cfg=dict(type='Maxpool2d', kernel_size=2, stride=(2, 2)), stages=(True, True, False, False), position='before_stage'), dict( cfg=dict(type='Maxpool2d', kernel_size=(2, 1), stride=(2, 1)), stages=(False, False, True, False), position='before_stage'), dict( cfg=dict( type='ConvModule', kernel_size=3, stride=1, padding=1, norm_cfg=dict(type='BN'), act_cfg=dict(type='ReLU')), stages=(True, True, True, True), position='after_stage') ])
Migration Guide - Dataset Annotation Loader¶
The annotation loaders, LmdbLoader
and HardDiskLoader
, are unified into AnnFileLoader
for a more consistent design and wider support on different file formats and storage backends. AnnFileLoader
can load the annotations from disk
(default), http
and petrel
backend, and parse the annotation in txt
or lmdb
format. LmdbLoader
and HardDiskLoader
are deprecated, and users are recommended to modify their configs to use the new AnnFileLoader
. Users can migrate their legacy loader HardDiskLoader
referring to the following example:
# Legacy config
train = dict(
type='OCRDataset',
...
loader=dict(
type='HardDiskLoader',
...))
# Suggested config
train = dict(
type='OCRDataset',
...
loader=dict(
type='AnnFileLoader',
file_storage_backend='disk',
file_format='txt',
...))
Similarly, using AnnFileLoader
with file_format='lmdb'
instead of LmdbLoader
is strongly recommended.
New Features & Enhancements¶
Update mmcv install by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/775
Upgrade isort by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/771
Automatically infer device for inference if not speicifed by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/781
Add open-mmlab precommit hooks by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/787
Add windows CI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/790
Add CurvedSyntext150k Converter by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/719
Add FUNSD Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/808
Support loading annotation file with petrel/http backend by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/793
Support different seeds on different ranks by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/820
Support json in recognition converter by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/844
Add args and docs for multi-machine training/testing by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/849
Add warning info for LineStrParser by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/850
Deploy openmmlab-bot by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/876
Add Tesserocr Inference by @garvan2021 in https://github.com/open-mmlab/mmocr/pull/814
Add LV Dataset Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/871
Add SROIE Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/810
Add NAF Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/815
Add DeText Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/818
Add IMGUR Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/825
Add ILST Converter by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/833
Add KAIST Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/835
Add IC11 (Born-digital Images) Data Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/857
Add IC13 (Focused Scene Text) Data Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/861
Add BID Converter by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/862
Add Vintext Converter by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/864
Add MTWI Data Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/867
Add COCO Text v2 Data Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/872
Add ReCTS Data Converter by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/892
Refactor ResNets by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/809
Bug Fixes¶
Bump mmdet version to 2.20.0 in Dockerfile by @GPhilo in https://github.com/open-mmlab/mmocr/pull/763
Update mmdet version limit by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/773
Minimum version requirement of albumentations by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/769
Disable worker in the dataloader of gpu unit test by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/780
Standardize the type of torch.device in ocr.py by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/800
Use RECOGNIZER instead of DETECTORS by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/685
Add num_classes to configs of ABINet by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/805
Support loading space character from dict file by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/854
Description in tools/data/utils/txt2lmdb.py by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/870
ignore_index in SARLoss by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/869
Fix a bug that may cause inplace operation error by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/884
Use hyphen instead of underscores in script args by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/890
Docs¶
Add deprecation message for deploy tools by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/801
Reorganizing OpenMMLab projects in readme by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/806
Add demo/README_zh.md by @EighteenSprings in https://github.com/open-mmlab/mmocr/pull/802
Add detailed version requirement table by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/778
Correct misleading section title in training.md by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/819
Update README_zh-CN document URL by @BeyondYourself in https://github.com/open-mmlab/mmocr/pull/823
translate testing.md. by @yangrisheng in https://github.com/open-mmlab/mmocr/pull/822
Fix confused description for load-from and resume-from by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/842
Add documents getting_started in docs/zh by @BeyondYourself in https://github.com/open-mmlab/mmocr/pull/841
Add the model serving translation document by @BeyondYourself in https://github.com/open-mmlab/mmocr/pull/845
Update docs about installation on Windows by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/852
Update tutorial notebook by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/853
Update Instructions for New Data Converters by @xinke-wang in https://github.com/open-mmlab/mmocr/pull/900
Brief installation instruction in README by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/897
update doc for ILST, VinText, BID by @Mountchicken in https://github.com/open-mmlab/mmocr/pull/902
Fix typos in readme by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/903
Recog dataset doc by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/893
Reorganize the directory structure section in det.md by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/894
New Contributors¶
@GPhilo made their first contribution in https://github.com/open-mmlab/mmocr/pull/763
@xinke-wang made their first contribution in https://github.com/open-mmlab/mmocr/pull/801
@EighteenSprings made their first contribution in https://github.com/open-mmlab/mmocr/pull/802
@BeyondYourself made their first contribution in https://github.com/open-mmlab/mmocr/pull/823
@yangrisheng made their first contribution in https://github.com/open-mmlab/mmocr/pull/822
@Mountchicken made their first contribution in https://github.com/open-mmlab/mmocr/pull/844
@garvan2021 made their first contribution in https://github.com/open-mmlab/mmocr/pull/814
Full Changelog: https://github.com/open-mmlab/mmocr/compare/v0.4.1…v0.5.0
v0.4.1 (27/01/2022)¶
Highlights¶
Visualizing edge weights in OpenSet KIE is now supported! https://github.com/open-mmlab/mmocr/pull/677
Some configurations have been optimized to significantly speed up the training and testing processes! Don’t worry - you can still tune these parameters in case these modifications do not work. https://github.com/open-mmlab/mmocr/pull/757
Now you can use CPU to train/debug your model! https://github.com/open-mmlab/mmocr/pull/752
We have fixed a severe bug that causes users unable to call
mmocr.apis.test
with our pre-built wheels. https://github.com/open-mmlab/mmocr/pull/667
New Features & Enhancements¶
Show edge score for openset kie by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/677
Download flake8 from github as pre-commit hooks by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/695
Deprecate the support for ‘python setup.py test’ by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/722
Disable multi-processing feature of cv2 to speed up data loading by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/721
Extend ctw1500 converter to support text fields by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/729
Extend totaltext converter to support text fields by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/728
Speed up training by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/739
Add setup multi-processing both in train and test.py by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/757
Support CPU training/testing by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/752
Support specify gpu for testing and training with gpu-id instead of gpu-ids and gpus by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/756
Remove unnecessary custom_import from test.py by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/758
Bug Fixes¶
Fix satrn onnxruntime test by @AllentDan in https://github.com/open-mmlab/mmocr/pull/679
Support both ConcatDataset and UniformConcatDataset by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/675
Fix bugs of show_results in single_gpu_test by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/667
Fix a bug for sar decoder when bi-rnn is used by @MhLiao in https://github.com/open-mmlab/mmocr/pull/690
Fix opencv version to avoid some bugs by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/694
Fix py39 ci error by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/707
Update visualize.py by @TommyZihao in https://github.com/open-mmlab/mmocr/pull/715
Fix link of config by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/726
Use yaml.safe_load instead of load by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/753
Add necessary keys to test_pipelines to enable test-time visualization by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/754
Docs¶
Fix recog.md by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/674
Add config tutorial by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/683
Add MMSelfSup/MMRazor/MMDeploy in readme by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/692
Add recog & det model summary by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/693
Update docs link by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/710
add pull request template.md by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/711
Add website links to readme by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/731
update readme according to standard by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/742
New Contributors¶
@MhLiao made their first contribution in https://github.com/open-mmlab/mmocr/pull/690
@TommyZihao made their first contribution in https://github.com/open-mmlab/mmocr/pull/715
Full Changelog: https://github.com/open-mmlab/mmocr/compare/v0.4.0…v0.4.1
v0.4.0 (15/12/2021)¶
Highlights¶
We release a new text recognition model - ABINet (CVPR 2021, Oral). With it dedicated model design and useful data augmentation transforms, ABINet can achieve the best performance on irregular text recognition tasks. Check it out!
We are also working hard to fulfill the requests from our community. OpenSet KIE is one of the achievement, which extends the application of SDMGR from text node classification to node-pair relation extraction. We also provide a demo script to convert WildReceipt to open set domain, though it cannot take the full advantage of OpenSet format. For more information, please read our tutorial.
APIs of models can be exposed through TorchServe. Docs
Breaking Changes & Migration Guide¶
Postprocessor¶
Some refactoring processes are still going on. For all text detection models, we unified their decode
implementations into a new module category, POSTPROCESSOR
, which is responsible for decoding different raw outputs into boundary instances. In all text detection configs, the text_repr_type
argument in bbox_head
is deprecated and will be removed in the future release.
Migration Guide: Find a similar line from detection model’s config:
text_repr_type=xxx,
And replace it with
postprocessor=dict(type='{MODEL_NAME}Postprocessor', text_repr_type=xxx)),
Take a snippet of PANet’s config as an example. Before the change, its config for bbox_head
looks like:
bbox_head=dict(
type='PANHead',
text_repr_type='poly',
in_channels=[128, 128, 128, 128],
out_channels=6,
loss=dict(type='PANLoss')),
Afterwards:
bbox_head=dict(
type='PANHead',
in_channels=[128, 128, 128, 128],
out_channels=6,
loss=dict(type='PANLoss'),
postprocessor=dict(type='PANPostprocessor', text_repr_type='poly')),
There are other postprocessors and each takes different arguments. Interested users can find their interfaces or implementations in mmocr/models/textdet/postprocess
or through our api docs.
New Config Structure¶
We reorganized the configs/
directory by extracting reusable sections into configs/_base_
. Now the directory tree of configs/_base_
is organized as follows:
_base_
├── det_datasets
├── det_models
├── det_pipelines
├── recog_datasets
├── recog_models
├── recog_pipelines
└── schedules
Most of model configs are making full use of base configs now, which makes the overall structural clearer and facilitates fair comparison across models. Despite the seemingly significant hierarchical difference, these changes would not break the backward compatibility as the names of model configs remain the same.
New Features¶
Support openset kie by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/498
Add converter for the Open Images v5 text annotations by Krylov et al. by @baudm in https://github.com/open-mmlab/mmocr/pull/497
Support Chinese for kie show result by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/464
Add TorchServe support for text detection and recognition by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/522
Save filename in text detection test results by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/570
Add codespell pre-commit hook and fix typos by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/520
Avoid duplicate placeholder docs in CN by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/582
Save results to json file for kie. by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/589
Add SAR_CN to ocr.py by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/579
mim extension for windows by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/641
Support muitiple pipelines for different datasets by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/657
ABINet Framework by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/651
Refactoring¶
Refactor textrecog config structure by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/617
Refactor text detection config by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/626
refactor transformer modules by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/618
refactor textdet postprocess by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/640
Docs¶
C++ example section by @apiaccess21 in https://github.com/open-mmlab/mmocr/pull/593
install.md Chinese section by @A465539338 in https://github.com/open-mmlab/mmocr/pull/364
Add Chinese Translation of deployment.md. by @fatfishZhao in https://github.com/open-mmlab/mmocr/pull/506
Fix a model link and add the metafile for SATRN by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/473
Improve docs style by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/474
Enhancement & sync Chinese docs by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/492
TorchServe docs by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/539
Update docs menu by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/564
Docs for KIE CloseSet & OpenSet by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/573
Fix broken links by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/576
Docstring for text recognition models by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/562
Add MMFlow & MIM by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/597
Add MMFewShot by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/621
Update model readme by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/604
Add input size check to model_inference by @mpena-vina in https://github.com/open-mmlab/mmocr/pull/633
Docstring for textdet models by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/561
Add MMHuman3D in readme by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/644
Use shared menu from theme instead by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/655
Refactor docs structure by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/662
Docs fix by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/664
Enhancements¶
Use bounding box around polygon instead of within polygon by @alexander-soare in https://github.com/open-mmlab/mmocr/pull/469
Add CITATION.cff by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/476
Add py3.9 CI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/475
update model-index.yml by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/484
Use container in CI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/502
CircleCI Setup by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/611
Remove unnecessary custom_import from train.py by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/603
Change the upper version of mmcv to 1.5.0 by @zhouzaida in https://github.com/open-mmlab/mmocr/pull/628
Update CircleCI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/631
Pass custom_hooks to MMCV by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/609
Skip CI when some specific files were changed by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/642
Add markdown linter in pre-commit hook by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/643
Use shape from loaded image by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/652
Cancel previous runs that are not completed by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/666
Bug Fixes¶
Modify algorithm “sar” weights path in metafile by @ShoupingShan in https://github.com/open-mmlab/mmocr/pull/581
Fix Cuda CI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/472
Fix image export in test.py for KIE models by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/486
Allow invalid polygons in intersection and union by default by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/471
Update checkpoints’ links for SATRN by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/518
Fix converting to onnx bug because of changing key from img_shape to resize_shape by @Harold-lkk in https://github.com/open-mmlab/mmocr/pull/523
Fix PyTorch 1.6 incompatible checkpoints by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/540
Fix paper field in metafiles by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/550
Unify recognition task names in metafiles by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/548
Fix py3.9 CI by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/563
Always map location to cpu when loading checkpoint by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/567
Fix wrong model builder in recog_test_imgs by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/574
Improve dbnet r50 by fixing img std by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/578
Fix resource warning: unclosed file by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/577
Fix bug that same start_point for different texts in draw_texts_by_pil by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/587
Keep original texts for kie by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/588
Fix random seed by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/600
Fix DBNet_r50 config by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/625
Change SBC case to DBC case by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/632
Fix kie demo by @innerlee in https://github.com/open-mmlab/mmocr/pull/610
fix type check by @cuhk-hbsun in https://github.com/open-mmlab/mmocr/pull/650
Remove depreciated image validator in totaltext converter by @gaotongxiao in https://github.com/open-mmlab/mmocr/pull/661
Fix change locals() dict by @Fei-Wang in https://github.com/open-mmlab/mmocr/pull/663
fix #614: textsnake targets by @HolyCrap96 in https://github.com/open-mmlab/mmocr/pull/660
New Contributors¶
@alexander-soare made their first contribution in https://github.com/open-mmlab/mmocr/pull/469
@A465539338 made their first contribution in https://github.com/open-mmlab/mmocr/pull/364
@fatfishZhao made their first contribution in https://github.com/open-mmlab/mmocr/pull/506
@baudm made their first contribution in https://github.com/open-mmlab/mmocr/pull/497
@ShoupingShan made their first contribution in https://github.com/open-mmlab/mmocr/pull/581
@apiaccess21 made their first contribution in https://github.com/open-mmlab/mmocr/pull/593
@zhouzaida made their first contribution in https://github.com/open-mmlab/mmocr/pull/628
@mpena-vina made their first contribution in https://github.com/open-mmlab/mmocr/pull/633
@Fei-Wang made their first contribution in https://github.com/open-mmlab/mmocr/pull/663
Full Changelog: https://github.com/open-mmlab/mmocr/compare/v0.3.0…0.4.0
v0.3.0 (25/8/2021)¶
Highlights¶
We add a new text recognition model – SATRN! Its pretrained checkpoint achieves the best performance over other provided text recognition models. A lighter version of SATRN is also released which can obtain ~98% of the performance of the original model with only 45 MB in size. (@2793145003) #405
Improve the demo script,
ocr.py
, which supports applying end-to-end text detection, text recognition and key information extraction models on images with easy-to-use commands. Users can find its full documentation in the demo section. (@samayala22, @manjrekarom) #371, #386, #400, #374, #428Our documentation is reorganized into a clearer structure. More useful contents are on the way! #409, #454
The requirement of
Polygon3
is removed since this project is no longer maintained or distributed. We unified all its references to equivalent substitutions inshapely
instead. #448
Breaking Changes & Migration Guide¶
Upgrade version requirement of MMDetection to 2.14.0 to avoid bugs #382
MMOCR now has its own model and layer registries inherited from MMDetection’s or MMCV’s counterparts. (#436) The modified hierarchical structure of the model registries are now organized as follows.
mmcv.MODELS -> mmdet.BACKBONES -> BACKBONES
mmcv.MODELS -> mmdet.NECKS -> NECKS
mmcv.MODELS -> mmdet.ROI_EXTRACTORS -> ROI_EXTRACTORS
mmcv.MODELS -> mmdet.HEADS -> HEADS
mmcv.MODELS -> mmdet.LOSSES -> LOSSES
mmcv.MODELS -> mmdet.DETECTORS -> DETECTORS
mmcv.ACTIVATION_LAYERS -> ACTIVATION_LAYERS
mmcv.UPSAMPLE_LAYERS -> UPSAMPLE_LAYERS
To migrate your old implementation to our new backend, you need to change the import path of any registries and their corresponding builder functions (including build_detectors
) from mmdet.models.builder
to mmocr.models.builder
. If you have referred to any model or layer of MMDetection or MMCV in your model config, you need to add mmdet.
or mmcv.
prefix to its name to inform the model builder of the right namespace to work on.
Interested users may check out MMCV’s tutorial on Registry for in-depth explanations on its mechanism.
New Features¶
Bug Fixes¶
Remove depreciated key in kie_test_imgs.py #381
Fix dimension mismatch in batch testing/inference of DBNet #383
Fix the problem of dice loss which stays at 1 with an empty target given #408
Fix undesired assignment to “pretrained” in test.py #418
Skip invalid annotations in totaltext_converter #438
Add zero division handler in poly utils, remove Polygon3 #448
Improvements¶
Replace lanms-proper with lanms-neo to support installation on Windows (with special thanks to @gen-ko who has re-distributed this package!)
Support MIM #394
Add tests for PyTorch 1.9 in CI #401
Enables fullscreen layout in readthedocs #413
General documentation enhancement #395
Update version checker #427
Add copyright info #439
Update citation information #440
Contributors¶
We thank @2793145003, @samayala22, @manjrekarom, @naarkhoo, @gen-ko, @duanjiaqi, @gaotongxiao, @cuhk-hbsun, @innerlee, @wdsd641417025 for their contribution to this release!
v0.2.1 (20/7/2021)¶
Highlights¶
New Features¶
Bug Fixes¶
Fix improper class ignorance in SDMGR Loss #221
Fix potential numerical zero division error in DRRG #224
Fix installing requirements with pip and mim #242
Fix dynamic input error of DBNet #269
Fix space parsing error in LineStrParser #285
Fix textsnake decode error #264
Correct isort setup #288
Fix a bug in SDMGR config #316
Fix kie_test_img for KIE nonvisual #319
Fix metafiles #342
Fix different device problem in FCENet #334
Ignore improper tailing empty characters in annotation files #358
Docs fixes #247, #255, #265, #267, #268, #270, #276, #287, #330, #355, #367
Improvements¶
Add backend for resizeocr #244
Skip image processing pipelines in SDMGR novisual #260
Speedup DBNet #263
Update mmcv installation method in workflow #323
Add support for ConcatDataset with two workflows #348
Add list_from_file and list_to_file utils #226
Speed up sort_vertex #239
Support distributed evaluation of KIE #234
Add pretrained FCENet on IC15 #258
Support CPU for OCR demo #227
Avoid extra image pre-processing steps #375
v0.2.0 (18/5/2021)¶
Highlights¶
Add the NER approach Bert-softmax (NAACL’2019)
Add the text detection method DRRG (CVPR’2020)
Add the text detection method FCENet (CVPR’2021)
Increase the ease of use via adding text detection and recognition end-to-end demo, and colab online demo.
Simplify the installation.
New Features¶
Bug Fixes¶
v0.1.0 (7/4/2021)¶
Highlights¶
MMOCR is released.
Main Features¶
Support text detection, text recognition and the corresponding downstream tasks such as key information extraction.
For text detection, support both single-step (
PSENet
,PANet
,DBNet
,TextSnake
) and two-step (MaskRCNN
) methods.For text recognition, support CTC-loss based method
CRNN
; Encoder-decoder (with attention) based methodsSAR
,Robustscanner
; Segmentation based methodSegOCR
; Transformer based methodNRTR
.For key information extraction, support GCN based method
SDMG-R
.Provide checkpoints and log files for all of the methods above.
mmocr.apis¶
mmocr.utils¶
mmocr.models¶
Common Backbones¶
Text Detection Detectors¶
Text Detection Heads¶
Text Detection Necks¶
Text Detection Losses¶
Text Detection Postprocessors¶
Text Recognition Recognizer¶
Text Recognition Backbones¶
Text Recognition Necks¶
Text Recognition Heads¶
Text Recognition Preprocessors¶
Text Recognition Backbones¶
Text Recognition Layers¶
Text Recognition Convertors¶
Text Recognition Encoders¶
Text Recognition Decoders¶
Text Recognition Fusers¶
Text Recognition Losses¶
KIE Extractors¶
KIE Heads¶
KIE Losses¶
NER Encoders¶
NER Decoders¶
NER Losses¶
Welcome to the OpenMMLab community¶
Scan the QR code below to follow the OpenMMLab team’s Zhihu Official Account and join the OpenMMLab team’s QQ Group, or join the official communication WeChat group by adding the WeChat, or join our Slack



We will provide you with the OpenMMLab community
📢 share the latest core technologies of AI frameworks
💻 Explaining PyTorch common module source Code
📰 News related to the release of OpenMMLab
🚀 Introduction of cutting-edge algorithms developed by OpenMMLab 🏃 Get the more efficient answer and feedback
🔥 Provide a platform for communication with developers from all walks of life
The OpenMMLab community looks forward to your participation! 👬