Research map / Vision-language and multimodal / Human and embodied modeling
Toward Understanding WordArt: Corner-Guided Transformer for Scene Text Recognition
Xudong Xie, Ling Fu, Zhifei Zhang, Zhaowen Wang, Xiang Bai
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 |
|---|---|---|---|
| add corner-map branch with corner-query cross-attention (ablation) | SATRN baseline · WordArt | word accuracy: 67.0 -> 69.1 | Improved |
| feed image instead of corner map to the extra branch (ablation) | SATRN baseline with cross-attention · WordArt | word accuracy: 67.0 -> 66.0 | Hurt |
| replace cross-attention with extra self-attention (no corner input) (ablation) | SATRN baseline · WordArt | word accuracy: 67.0 -> 67.6 | No clear effect |
| use Harris instead of Shi-Tomasi corner detector (ablation) | Corner-guided encoder · WordArt | word accuracy: 69.1 -> 68.4 | Hurt |
| use SuperPoint instead of Shi-Tomasi corner detector (ablation) | Corner-guided encoder · WordArt | word accuracy: 69.1 -> 69.0 | No clear effect |
| use corner as key/value instead of query (ablation) | Corner-guided encoder · WordArt | word accuracy: 69.1 -> 66.9 | Hurt |
| fuse corner and image by concatenation instead of corner-query attention (ablation) | SATRN baseline · WordArt | word accuracy: 69.1 -> 67.0 | Hurt |
| fuse corner and image by addition instead of corner-query attention (ablation) | SATRN baseline · WordArt | word accuracy: 69.1 -> 66.6 | Hurt |
| fuse corner and image by multiplication instead of corner-query attention (ablation) | SATRN baseline · WordArt | word accuracy: 69.1 -> 67.4 | Hurt |
| add character contrastive loss (lambda 0.1, tau 0.1, d 2048) (ablation) | SATRN baseline · WordArt | word accuracy: 67.0 -> 68.6 | Improved |
| lower temperature tau to 0.05 in CC loss (ablation) | CC loss (lambda 0.1, d 512) · WordArt | word accuracy: 68.1 -> 66.5 | Hurt |
| raise CC loss weight lambda to 1 (ablation) | CC loss (tau 0.1, d 2048) · WordArt | word accuracy: 68.6 -> 66.6 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RobustScanner — RobustScanner: Dynamically Enhancing Positional Clues for Robust Text Recognition (ECCV 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods (ECCV 2026 · similarity 0.90)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.87)
- WAS: Dataset and Methods for Artistic Text Segmentation (ECCV 2024 · similarity 0.87)
- Character Region Awareness for Text Detection (CVPR 2019 · similarity 0.87)
- Text Image Inpainting via Global Structure-Guided Diffusion Models (AAAI 2024 · similarity 0.86)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.86)
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.
