- Task 1 - Text Localization
- Task 2 - Script identification
- Task 3 - Joint text detection and script identification
- Task 4 - End-to-End text detection and recognition
method: SituTech_OCR2021-03-11
Authors: Kui Lyu, Chuanhe Liu
Affiliation: Beijing Situ Vision Technologies Co. Ltd
Email: lvkui@situdata.com
Description: In this work, we design an elegant text detection model. Our detector is similar to DBNet, but there are some difference. More specifically, we have introduced an advanced detector backbone, a classic network EfficientDet, with flexible scales and stronger ability to extract features. Another breakthrough is that we optimized the label generation strategy in DBNet. In the original work, the positive area generation and the expansion of the positive area to the bounding box used the Vatti clipping algorithm, which is less robust with different area perimeter ratios. We optimized this function to make the label transform between positive area and bounding box more reasonable.
If you have any questions, please contact us.
SituAIgorithm Team, Beijing Situ Vision Technologies Co. Ltd
method: multi-stage_text_detector_v42019-06-04
Authors: Pengfei Wang~*, Mengyi En*, Xiaoqiang Zhang*, Chengquan Zhang*
Affiliation: VIS-VAR Team, Baidu Inc.*; Xidian University~
Description: The method mainly relies on a two-stage text detector, namely LOMO [1], which is inspired by Mask-R-CNN and where an iterative refinement module is introduced to refine the boundary of text region once or more times during testing to get the more accurate detection results. As extra data sets, ICDAR15 and partial KAIST are also used in the training phase. Multi-scale testing is adopted and the final result is boosted from LOMOs with Resnet-50 and Inception-v4 as different backbones.
*This work is done when Pengfei Wang is an intern at Baidu Inc.
method: PMTD2019-05-30
Authors: Xuebo Liu，Jingchao Liu, Ding Liang
Description: Pyramid Mask Text Detector predicts a soft mask for every text instance, and then uses the plane clustering algorithm to get the final text box. See https://arxiv.org/abs/1903.11800 for detail. Trained model and inference code will be released. If you have questions, please feel free to contact Jingchao Liu (liujingchao@sensetime.com) and Xuebo Liu (liuxuebo@sensetime.com)
| Date | Method | Hmean | Precision | Recall | Average Precision | |||
|---|---|---|---|---|---|---|---|---|
| 2021-03-11 | SituTech_OCR | 84.80% | 93.41% | 77.64% | 72.43% | |||
| 2019-06-04 | multi-stage_text_detector_v4 | 83.59% | 87.75% | 79.80% | 70.00% | |||
| 2019-05-30 | PMTD | 82.53% | 87.47% | 78.12% | 75.80% | |||
| 2019-06-02 | A two-stage text detector based on cascade rcnn(using total 10000 images of mlt19) | 78.38% | 82.26% | 74.85% | 71.27% | |||
| 2019-05-31 | A two-stage text detector based on cascade rcnn | 78.11% | 82.89% | 73.85% | 70.31% | |||
| 2019-05-26 | two stage text detector | 75.04% | 82.61% | 68.74% | 65.29% | |||
| 2019-06-03 | CRAFTS | 70.86% | 81.42% | 62.73% | 56.63% | |||
| 2019-05-28 | CRAFTS(Initial) | 68.11% | 79.51% | 59.56% | 54.50% | |||
| 2019-05-24 | PSENet_v1 | 65.83% | 73.52% | 59.59% | 52.73% |