Research map / Label-efficient and robust learning / Self-supervised and contrastive
Appearance Discrepancy-guided Sequence Hybrid Masking for Robust Scene Text Recognition
Shihao Zou, Wei Wei, Leyang Xu, Kaihe Xu, Wenfeng Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use sequence hybrid masking instead of random masking (ablation) | DSHM pre-training · Challenge sets avg (HOST, WOST, ArT, COCO, Uber), 5M pre-training images | avg word accuracy: 75.5 -> 79.0 | Improved |
| add discrepancy guidance to random masking (HM vs RM) (ablation) | DSHM pre-training · Challenge sets avg (HOST, WOST, ArT, COCO, Uber), 5M pre-training images | avg word accuracy: 75.5 -> 77.2 | Improved |
| add sequence awareness (block expansion) to hybrid masking (ablation) | DSHM pre-training · Challenge sets avg (HOST, WOST, ArT, COCO, Uber), 5M pre-training images | avg word accuracy: 77.2 -> 79.0 | Improved |
| add discrepancy-weighted reconstruction loss (ablation) | DSHM pre-training · Challenge sets avg (HOST, WOST, ArT, COCO, Uber), 5M pre-training images | avg word accuracy: 77.3 -> 79.0 | Improved |
| add CLS consistency loss (ablation) | DSHM pre-training · Challenge sets avg (HOST, WOST, ArT, COCO, Uber), 5M pre-training images | avg word accuracy: 78.3 -> 79.0 | Improved |
| add discrepancy-conditioned token in decoder (ablation) | DSHM pre-training · Challenge sets avg (HOST, WOST, ArT, COCO, Uber), 5M pre-training images | avg word accuracy: 76.6 -> 79.0 | Improved |
| hard/random mask ratio 0.5/0.3 instead of 0.6/0.2 (ablation) | DSHM pre-training · Challenge sets avg (HOST, WOST, ArT, COCO, Uber), 5M pre-training images | avg word accuracy: 77.3 -> 79.0 | Improved |
| hard/random mask ratio 0.5/0.3 instead of 0.4/0.4 (ablation) | DSHM pre-training · Challenge sets avg (HOST, WOST, ArT, COCO, Uber), 5M pre-training images | avg word accuracy: 76.6 -> 79.0 | Improved |
| use DSHM pre-training instead of LMIM | LMIM · Union14M benchmark (avg) | word accuracy (%): 86.3 -> 87.0 | Improved |
| use DSHM pre-training instead of LMIM | LMIM · Chinese benchmark (avg, with pre-training) | accuracy (%): 82.2 -> 83.0 | Improved |
| use DSHM pre-training instead of LMIM | LMIM · HOST (occluded/challenging) | word accuracy (%): 69.6 -> 72.8 | Improved |
Closest papers by meaning
- Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition (CVPR 2025 · similarity 0.89)
- SyncMask: Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining (CVPR 2024 · similarity 0.85)
- Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection (CVPR 2025 · similarity 0.85)
- T-MARS: Improving Visual Representations by Circumventing Text Feature Learning (ICLR 2024 · similarity 0.85)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.85)
- On the Role of Discrete Tokenization in Visual Representation Learning (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
