liviust / TCM

Turning a CLIP Model into a Scene Text Detector (CVPR2023)

Turning a CLIP Model into a Scene Text Detector

This repository is build upon mmocr 0.4.0.

NightTime-ArT Dataset

NightTime-ArT dataset, collected from ArT, can be downloaded from here.

Usage

Environment

• cuda 11.1
• torch=1.8.0
• torchvision=0.9.0
• timm=0.4.12
• mmcv-full=1.3.17
• mmseg=0.20.2
• mmdet=2.19.1
• mmocr=0.4.0

The code is based on mmocr. Please first install the mmcv-full and mmocr following the official guidelines (mmocr).

Dataset

• Please following the mmocr official guidelines to prepare the datasets accordingly.

• Configure the dataset path in ocrclip/configs/_base_/det_datasets.

Pre-trained CLIP Models

• Download the pre-trained CLIP models (RN50.pt) and save them to the pretrained folder.
• Configure the pre-trained CLIP models path in config file as

model = dict(
    pretrained='xxx/ocrclip/pretrained/RN50.pt',
    )

Pretraining & Training & Evaluation

To pretrain the TCM model on SynthText/Synth150k, please configure the corresponding dataset path, then run:

bash dist_train.sh configs/textdet/xxnet/xxx.py 8

To finetune the TCM model based on pretrained model, please configure the load_from to the pretrained checkpoint path, then run:

bash dist_train.sh configs/textdet/xxnet/xxx.py 8

To evaluate the performance with checkpoint, run:

bash dist_test.sh configs/textdet/xxnet/xxx.py /path/to/checkpoint 1 --eval hmean-iou

Results

Method	Data	F-measure	Model
TCM-DB	TD	88.8%	config weights
TCM-DB	IC15	88.8%	config weights
TCM-DB	CTW	85.1%	config
TCM-DB	TT	85.9%	config

TODO

• Add FastTCM
• Migration from mmocr 0.4.0 to mmocr 1.0.0
• Refactor and clean code
• Release domain adaptation setting

Cites

If you find this project helpful for your research, please consider citing the paper

@inproceedings{Yu2023TurningAC,
  title={Turning a CLIP Model into a Scene Text Detector},
  author={Wenwen Yu and Yuliang Liu and Wei Hua and Deqiang Jiang and Bo Ren and Xiang Bai},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2023}
}

Licence

This project is under the CC-BY-NC 4.0 license. See LICENSE for more details.

Acknowledges

The project partially based on MMOCR, CLIP, DenseCLIP. Thanks for their great works.