TransDETR: End-to-end Video Text Spotting with Transformer

Introduction

[End-to-end Video Text Spotting with Transformer]

Video text spotting(VTS) is the task that requires simultaneously detecting, tracking and recognizing text instances in video. Recent methods typically develop sophisticated pipelines based on Intersection over Union (IoU) or appearance similarity in adjacent frames to tackle this task. In this paper, rooted in Transformer sequence modeling, we propose a novel video text DEtection, Tracking, and Recognition framework (TransDETR), which views the VTS task as a direct long-sequence temporal modeling problem.

Link to our new benchmark BOVText: A Large-Scale, Bilingual Open World Dataset for Video Text Spotting

Updates

• (08/04/2021) Refactoring the code.

• (10/20/2021) The complete code has been released .

Performance

ICDAR2015(video) Tracking challenge

Methods	MOTA	MOTP	IDF1	Mostly Matched	Partially Matched	Mostly Lost
TransDETR	47.5	74.2	65.5	832	484	600

Models are also available in Baidu Drive (password:4n03) or Google Drive.

ICDAR2015(video) Video Text Spotting challenge

Methods	MOTA	MOTP	IDF1	Mostly Matched	Partially Matched	Mostly Lost
TransDETR	58.4	75.2	70.4	614	326	427

Notes

• The training time is on 8 NVIDIA V100 GPUs with batchsize 16.
• We use the models pre-trained on COCOTextV2.
• We do not release the recognition code due to the company's regulations.

Demo

Installation

The codebases are built on top of Deformable DETR and MOTR.

• Linux, CUDA>=9.2, GCC>=5.4

• Python>=3.7

  We recommend you to use Anaconda to create a conda environment:

  conda create -n TransDETR python=3.7 pip

  Then, activate the environment:

  conda activate TransDETR

• PyTorch>=1.5.1, torchvision>=0.6.1 (following instructions here)

  For example, if your CUDA version is 9.2, you could install pytorch and torchvision as following:

  conda install pytorch=1.5.1 torchvision=0.6.1 cudatoolkit=9.2 -c pytorch

• Other requirements

  pip install -r requirements.txt

• Build MultiScaleDeformableAttention and Rotated ROIAlign

  cd ./models/ops
  sh ./make.sh
  
  cd ./models/Rotated_ROIAlign
  python setup.py build_ext --inplace

Usage

Dataset preparation

1. Please download ICDAR2015 and COCOTextV2 dataset and organize them like FairMOT as following:

.
├── COCOText
│   ├── images
│   └── labels_with_ids
├── ICDAR15
│   ├── images
│       ├── track
│           ├── train
│           ├── val
│   ├── labels
│       ├── track
│           ├── train
│           ├── val

2. You also can use the following script to generate txt file:

cd tools/gen_labels
python3 gen_labels_COCOTextV2.py
python3 gen_labels_15.py
python3 gen_labels_YVT.py
cd ../../

Training and Evaluation

Training on single node

You can download COCOTextV2 pretrained weights from Pretrained TransDETR Baidu Drive (password:s93i) Google Drive. Or training by youself:

sh configs/r50_TransDETR_pretrain_COCOText.sh

Then training on ICDAR2015 with 8 GPUs as following:

sh configs/r50_TransDETR_train.sh

Evaluation on ICDAR13 and ICDAR15

You can download the pretrained model of TransDETR (the link is in "Main Results" session), then run following command to evaluate it on ICDAR2015 dataset:

sh configs/r50_TransDETR_eval.sh

Visualization

For visual in demo video, you can enable 'vis=True' in eval.py like:

--show

then run the script:

python tools/vis.py

License

TransDETR is released under MIT License.

Citing

If you use TransDETR in your research or wish to refer to the baseline results published here, please use the following BibTeX entries:

@article{wu2022transdetr,
  title={End-to-End Video Text Spotting with Transformer},
  author={Weijia Wu, Chunhua Shen, Yuanqiang Cai, Debing Zhang, Ying Fu, Ping Luo, Hong Zhou},
  journal={arxiv},
  year={2022}
}