This code repository contains the implementations of the paper Text Perceptron: Towards End-to-End Arbitrary-Shaped Text Spotting (AAAI 2020).
This repository only contains the detection part of Text Perceptron. For the entire end-to-end implementation, please going to Text Perceptron Spot
Original images can be downloaded from: Total-Text , SCUT-CTW1500.
The formatted training datalist and test datalist can be found in demo/text_detection/datalist/
Modified the paths of "imgs"/ "pretrained_model"/ "work_space" in the config files demo/text_detection/text_perceptron_det/config/tp_r50_3stages_enlarge.py.
Run the following bash command in the command line,
cd $DAVAR_LAB_OCR_ROOT$/demo/text_detection/text_perceptron_det/
bash dist_train.shWe provide the implementation of online validation. If you want to close it to save training time, you may modify the startup script to add
--no-validatecommand.
We provide a demo of forward inference and visualization. You can modify the paths (test_dataset, image_prefix, etc.) in the testing script, and start testing:
python test.py Some visualization of detection results are shown:
The offline evaluation tool can be found in davarocr/demo/text_detection/evaluation/.
All of the models are re-implemented and well trained in the based on the opensourced framework mmdetection. So, the results might be slightly different from reported results.
Results on various datasets and trained models download:
| Dataset | Backbone | Pretrained | Test Scale | Precision | Recall | Hmean | Links |
|---|---|---|---|---|---|---|---|
| Total-Text (Reported) | ResNet-50-3stages-enlarge | SynthText | L-1350 | 88.1 | 78.9 | 83.3 | - |
| Total-Text | ResNet-50-3stages-enlarge | SynthText | L-1350 | 89.0 | 81.1 | 84.8 | config, pth (Access Code: 2345) |
| Total-Text | ResNet-50 | SynthText+IC17, pth (Access Code: 8dn0) | L-1350 | 84.5 | 80.4 | 82.4 | config, pth (Access Code: p692) |
| SCUT-CTW1500 (Reported) | ResNet-50-3stages-enlarge | SynthText | L-1250 | 88.7 | 78.2 | 83.1 | - |
| SCUT-CTW1500 | ResNet-50-3stages-enlarge | SynthText | L-1250 | 86.2 | 79.9 | 82.9 | config, pth (Access Code: t2z9) |
| SCUT-CTW1500 (considers NOT CARE) | ResNet-50-3stages-enlarge | SynthText | L-1250 | 85.9 | 83.6 | 84.7 | config, pth (Access Code: t2z9) |
| SCUT-CTW1500 | ResNet-50 | SynthText+IC17, pth (Access Code: 8dn0) | L-1250 | 87.7 | 79.8 | 83.6 | config, pth(Access Code: duuo) |
| SCUT-CTW1500 (considers NOT CARE) | ResNet-50 | SynthText+IC17, pth (Access Code: 8dn0) | L-1250 | 87.5 | 84.5 | 85.9 | config, pth(Access Code: duuo) |
The original SCUT-CTW1500 was released to only support text detection task, where all text instances are considered. When this dataset was used in text spotting task, some instances (blurred or in other languages) were labeled as not care.
Note: Above links are from BaiduYunPan. Models can also be downloaded from Google Drive
If you find this repository is helpful to your research, please feel free to cite us:
@inproceedings{qiao2020text,
title={Text Perceptron: Towards End-to-End Arbitrary-Shaped Text Spotting},
author={Qiao, Liang and Tang, Sanli and Cheng, Zhanzhan and Xu, Yunlu and Niu, Yi and Pu, Shiliang and Wu, Fei},
booktitle={Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI)},
pages={11899-11907},
year={2020}
}This project is released under the Apache 2.0 license
If there is any suggestion and problem, please feel free to contact the author with qiaoliang6@hikvision.com or chengzhanzhan@hikvision.com.