Awesome

Introduction

vedastr is an open source scene text recognition toolbox based on PyTorch. It is designed to be flexible in order to support rapid implementation and evaluation for scene text recognition task.

Features

Modular design
We decompose the scene text recognition framework into different components and one can easily construct a customized scene text recognition framework by combining different modules.
Flexibility
vedastr is flexible enough to be able to easily change the components within a module.
Module expansibility
It is easy to integrate a new module into the vedastr project.
Support of multiple frameworks
The toolbox supports several popular scene text recognition framework, e.g., CRNN, TPS-ResNet-BiLSTM-Attention, Transformer, etc.
Good performance
We re-implement the best model in deep-text-recognition-benchmark and get better average accuracy. What's more, we implement a simple baseline(ResNet-FC) and the performance is acceptable.

License

This project is released under Apache 2.0 license.

Benchmark and model zoo

Note:

We use MJSynth(MJ) and SynthText(ST) as training data, and test the models on IIIT5K_3000, SVT, IC03_867, IC13_1015, IC15_2077, SVTP, CUTE80. You can find the datasets below.

MODEL	CASE SENSITIVE	IIIT5k_3000	SVT	IC03_867	IC13_1015	IC15_2077	SVTP	CUTE80	AVERAGE
ResNet-CTC	False	87.97	84.54	90.54	88.28	67.99	72.71	77.08	81.58
ResNet-FC	False	88.80	88.41	92.85	90.34	72.32	79.38	76.74	84.24
TPS-ResNet-BiLSTM-Attention	False	90.93	88.72	93.89	92.12	76.41	80.31	79.51	86.49
Small-SATRN	False	91.97	88.10	94.81	93.50	75.64	83.88	80.90	87.19

TPS : Spatial transformer network

Small-SATRN: On Recognizing Texts of Arbitrary Shapes with 2D Self-Attention, training phase is case sensitive while testing phase is case insensitive.

AVERAGE : Average accuracy over all test datasets

CASE SENSITIVE : If true, the output is case sensitive and contain common characters. If false, the output is not case sensetive and contains only numbers and letters.

Installation

Requirements

Linux
Python 3.6+
PyTorch 1.4.0 or higher
CUDA 9.0 or higher

We have tested the following versions of OS and softwares:

OS: Ubuntu 16.04.6 LTS
CUDA: 10.2
Python 3.6.9
PyTorch: 1.5.1

Install vedastr

Create a conda virtual environment and activate it.

conda create -n vedastr python=3.6 -y
conda activate vedastr

Install PyTorch and torchvision following the official instructions, e.g.,

conda install pytorch torchvision -c pytorch

Clone the vedastr repository.

git clone https://github.com/Media-Smart/vedastr.git
cd vedastr
vedastr_root=${PWD}

Install dependencies.

pip install -r requirements.txt

Prepare data

Download Lmdb data from deep-text-recognition-benchmark, which contains training, validation and evaluation data. Note: we use the ST dataset released by ASTER.
Make directory data as follows:

cd ${vedastr_root}
mkdir ${vedastr_root}/data

Put the download LMDB data into this data directory, the structure of data directory will look like as follows:

data
└── data_lmdb_release
    ├── evaluation
    ├── training
    │   ├── MJ
    │   │   ├── MJ_test
    │   │   ├── MJ_train
    │   │   └── MJ_valid
    │   └── ST
    └── validation

Train

Config

Modify configuration files in configs/ according to your needs(e.g. configs/tps_resnet_bilstm_attn.py).

# train using GPUs with gpu_id 0, 1, 2, 3
python tools/train.py configs/tps_resnet_bilstm_attn.py "0, 1, 2, 3"

Snapshots and logs by default will be generated at ${vedastr_root}/workdir/name_of_config_file(you can specify workdir in config files).

Test

Config

Modify configuration as you wish(e.g. configs/tps_resnet_bilstm_attn.py).

# test using GPUs with gpu_id 0, 1
./tools/dist_test.sh configs/tps_resnet_bilstm_attn.py path/to/checkpoint.pth "0, 1"

Inference

# inference using GPUs with gpu_id 0
python tools/inference.py configs/tps_resnet_bilstm_attn.py checkpoint_path img_path "0"

Deploy

Install volksdep following the official instructions
Benchmark (optional)

# Benchmark model using GPU with gpu_id 0
CUDA_VISIBLE_DEVICES="0" python tools/benchmark.py configs/resnet_ctc.py checkpoint_path out_path --dummy_input_shape "3,32,100"

More available arguments are detailed in tools/deploy/benchmark.py.

The result of resnet_ctc is as follows(test device: Jetson AGX Xavier, CUDA:10.2):

framework	version	input shape	data type	throughput(FPS)	latency(ms)
PyTorch	1.5.0	(1, 1, 32, 100)	fp32	64	15.81
TensorRT	7.1.0.16	(1, 1, 32, 100)	fp32	109	9.66
PyTorch	1.5.0	(1, 1, 32, 100)	fp16	113	10.75
TensorRT	7.1.0.16	(1, 1, 32, 100)	fp16	308	3.55
TensorRT	7.1.0.16	(1, 1, 32, 100)	int8(entropy_2)	449	2.38

Export model to ONNX format

# export model to onnx using GPU with gpu_id 0
CUDA_VISIBLE_DEVICES="0" python tools/torch2onnx.py configs/resnet_ctc.py checkpoint_path --dummy_input_shape "3,32,100" --dynamic_shape

More available arguments are detailed in tools/torch2onnx.py.

Inference SDK

You can refer to FlexInfer for details.

Citation

If you use this toolbox or benchmark in your research, please cite this project.

@misc{2020vedastr,
    title  = {vedastr: A Toolbox for Scene Text Recognition},
    author = {Sun, Jun and Cai, Hongxiang and Xiong, Yichao},
    url    = {https://github.com/Media-Smart/vedastr},
    year   = {2020}
}

Contact

This repository is currently maintained by Jun Sun(@ChaseMonsterAway), Hongxiang Cai (@hxcai), Yichao Xiong (@mileistone).