mxin262 / Bridging-Text-Spotting

(CVPR 2024) Bridging the Gap Between End-to-End and Two-Step Text Spotting.

Bridging the Gap Between End-to-End and Two-Step Text Spotting

Updates | Main Results | Usage | Citation | Acknowledgement

This is the official repo for the paper "Bridging the Gap Between End-to-End and Two-Step Text Spotting", which is accepted to CVPR 2024.

Updates

[Apr.08, 2024]The paper is submitted to ArXiv. 🔥🔥🔥 You can use only a single 3090 to train the model.

Main Results

Total-Text

Method	Det-P	Det-R	Det-F1	E2E-None	E2E-Full	Weights
DG-Bridge Spotter	92.0	86.5	89.2	83.3	88.3	Google Drive

CTW1500

Method	Det-P	Det-R	Det-F1	E2E-None	E2E-Full	Weights
DG-Bridge Spotter	92.1	86.2	89.0	69.8	83.9	Google Drive

ICDAR 2015 (IC15)

Backbone	Det-P	Det-R	Det-F1	E2E-S	E2E-W	E2E-G	Weights
TG-Bridge Spotter	93.8	87.5	90.5	89.1	84.2	80.4	Google Drive

Usage

It's recommended to configure the environment using Anaconda. Python 3.8 + PyTorch 1.9.1 (or 1.9.0) + CUDA 11.1 + Detectron2 (v0.6) are suggested.

• Installation

conda create -n Bridge python=3.8 -y
conda activate Bridge
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install opencv-python scipy timm shapely albumentations Polygon3
pip install setuptools==59.5.0
git clone https://github.com/mxin262/Bridging-Text-Spotting.git
cd Bridging-Text-Spotting
cd detectron2
python setup.py build develop
cd ..
python setup.py build develop

• Data Preparation

Total-Text (including rotated images): link

CTW1500 (including rotated images): link

ICDAR2015 (images): link.

Json files for Total-Text and CTW1500: OneDrive | BaiduNetdisk(44yt)

Json files for ICDAR2015: link,

Evaluation files: link. Extract them under datasets folder.

Organize them as follows:

|- datasets
   |- totaltext
   |  |- test_images_rotate
   |  |- train_images_rotate
   |  |- test_poly.json
   |  |- test_poly_rotate.json
   |  |─ train_poly_ori.json
   |  |─ train_poly_pos.json
   |  |─ train_poly_rotate_ori.json
   |  └─ train_poly_rotate_pos.json
   |- ctw1500
   |  |- test_images
   |  |- train_images_rotate
   |  |- test_poly.json
   |  └─ train_poly_rotate_pos.json
   |- icdar2015
   |   | - test_images
   |   | - train_images
   |   | - test.json
   |   | - train.json
   |- evaluation
   |  |- lexicons
   |  |- gt_totaltext.zip
   |  |- gt_ctw1500.zip
   |  |- gt_inversetext.zip
   |  └─ gt_totaltext_rotate.zip

The generation of positional label form for DPText-DETR is provided in process_positional_label.py

• Training

Download the pre-trained model from DPText-DETR, DiG, TESTR.

With the pre-trained model, use the following command to fine-tune it on the target benchmark. For example:

python tools/train_net.py --config-file configs/Bridge/TotalText/R_50_poly.yaml --num-gpus 4  MODEL.WEIGHTS totaltext_final.pth

• Evaluation

python tools/train_net.py --config-file ${CONFIG_FILE} --eval-only MODEL.WEIGHTS ${MODEL_PATH}

• Inference & Visualization

python demo/demo.py --config-file ${CONFIG_FILE} --input ${IMAGES_FOLDER_OR_ONE_IMAGE_PATH} --output ${OUTPUT_PATH} --opts MODEL.WEIGHTS <MODEL_PATH>

Citation

If you find Bridge Text Spotting useful in your research, please consider citing:

@inproceedings{huang2024bridge,
  title={Bridging the Gap Between End-to-End and Two-Step Text Spotting},
  author={Huang, Mingxin and Li, Hongliang and Liu, Yuliang and Bai, Xiang and Jin, Lianwen},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2024}
}

Acknowlegement

AdelaiDet, DPText-DETR, DiG, TESTR. Thanks for their great work!