- Task 1 - Detection
- Task 2 - Detection-Linking
- Task 3 - Detection-Recognition
- Task 4 - Detection-Recognition-Linking
method: MapText Detection-Recognition-Linking Strong Pipeline2024-05-06
Authors: Yu Xie, Jielei Zhang, Ziyue Wang, Yuchen He, Yihan Meng, Weihang Wang, Peiyi Li, Longwen Gao
Affiliation: Bilibili Inc.
Description: In the Detection-Recognition-Linking task of MapText, we used ViTAE-v2 to extract global features, utilizing an encoder-decoder network architecture (DeepSolo). Data augmentation techniques such as cropping, scaling, saturation, and contrast adjustment were applied. Pre-training was conducted using available real datasets (TextOCR, TotalText, IC15, MLT2017). The model was fine-tuned on the MapText dataset, and post-processing methods were employed.
Zhang, Q., Xu, Y., Zhang, J., & Tao, D. (2023). Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. International Journal of Computer Vision, 131(5), 1141-1162.
Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., & Tao, D. (2023). Deepsolo: Let transformer decoder with explicit points solo for text spotting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 19348-19357).
method: Baseline TESTR Checkpoint2024-03-26
Authors: Organizers
Affiliation: ICDAR'24 RRC-MapText
Description: TESTR checkpoint is used without any additional modifications or finetuning. The model checkpoint version with polygon prediction head and fine-tuned on TotalText was used. (Note that no links are predicted.)
Date | Method | Quality | Char Accuracy | F-score | Tightness | Recall | Precision | |||
---|---|---|---|---|---|---|---|---|---|---|
2024-05-06 | MapText Detection-Recognition-Linking Strong Pipeline | 17.08% | 55.73% | 44.57% | 68.76% | 44.99% | 44.17% | |||
2024-03-26 | Baseline TESTR Checkpoint | 6.05% | 41.46% | 21.22% | 68.73% | 12.95% | 58.71% |