Research map / Image generation / Text-to-image diffusion
Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors
Tongkun Guan, Wei Shen, Xue Yang, Xuehui Wang, Xiaokang Yang
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 standard student-teacher framework to DBNet pre-training (ablation) | DBNet · Total-Text (20% data) | F-measure: 84.7 -> 86.01 | Improved |
| add GlyphMix to student-teacher framework (ablation) | FreeReal · Total-Text (20% data) | F-measure: 86.01 -> 87.82 | Improved |
| add character region awareness (character-level boxes) (ablation) | FreeReal · Total-Text (20% data) | F-measure: 86.01 -> 86.43 | Improved |
| combine CRA and GlyphMix (ablation) | FreeReal · Total-Text (20% data) | F-measure: 87.82 -> 88.10 | Improved |
| add graffiti-like inter-domain mixing (GIM) (ablation) | GlyphMix · Total-Text (20% data) | F-measure: 86.43 -> 87.71 | Improved |
| add text-balanced intra-domain mixing (TIM) (ablation) | GlyphMix · Total-Text (20% data) | F-measure: 86.43 -> 87.65 | Improved |
| use GlyphMix instead of Mixup, CutMix or ClassMix for domain bridging (ablation) | FreeReal · Total-Text (20% data) | F-measure: 86.99 -> 88.10 vs ClassMix | Improved |
| use GlyphMix versus recent semi-supervised methods (ablation) | FreeReal · Total-Text (20% data) | F-measure: 87.22 -> 88.10 vs Unimatch | Improved |
| replace GlyphMix with GTA semi-supervised method (ablation) | FreeReal student-teacher · Total-Text (20% data) | F-measure: 86.01 -> 84.02 | Hurt |
| cluster per cropped text instance instead of whole image for glyph mask (ablation) | GlyphMix · TotalText glyph segmentation | glyph mask accuracy: 32.35% -> 80.43% | Improved |
| linearly decay gamma from 80 to 20 instead of fixing it (ablation) | FreeReal · Total-Text (20% data) | F-measure: 87.94 -> 88.10 | Improved |
| increase pre-training data fraction from 10% to 100% (ablation) | FreeReal · Total-Text (DBNet) | F-measure: 87.5 -> 88.9 | Improved |
Closest papers by meaning
- TextNeRF: A Novel Scene-Text Image Synthesis Method based on Neural Radiance Fields (CVPR 2024 · similarity 0.87)
- GlyphSR: A Simple Glyph-Aware Framework for Scene Text Image Super-Resolution (AAAI 2025 · similarity 0.87)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.87)
- GlyphMastero: A Glyph Encoder for High-Fidelity Scene Text Editing (CVPR 2025 · similarity 0.86)
- Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data (CVPR 2025 · similarity 0.86)
- Revisiting Tampered Scene Text Detection in the Era of Generative AI (AAAI 2025 · 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.
