Research map / Vision-language and multimodal / CLIP and zero-shot
Beyond Script Family Boundaries: Towards Unified Open-Set Scene Text Recognition
Chang Liu, Elisa Barney Smith
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 CAM-CTC alignment instead of position-wise attention alignment (ablation) | word recognition pipeline · Korean (GZSL) | ACR: 15.62 -> 19.18 | Improved |
| use CAM-CTC alignment instead of CTC squeeze alignment (ablation) | word recognition pipeline · Synthetic Yi (GZSL) | ACR: 6.68 -> 8.51 | Improved |
| use CAM-CTC alignment instead of position-wise alignment (ablation) | word recognition pipeline · Japanese (GZSL) | ACR: 41.59 -> 42.36 | No clear effect |
| add synthetic word co-training (ablation) | Ours-CJK · Synthetic Yi (GZSL) | ACR: 8.51 -> 11.42 | Improved |
| add synthetic character co-training (ablation) | Ours-CJK · Synthetic Yi (GZSL) | ACR: 8.51 -> 12.05 | Improved |
| add both word and character synthetic co-training (ablation) | Ours-CJK · Synthetic Yi (GZSL) | ACR: 8.51 -> 17.07 | Improved |
| add both word and character synthetic co-training (ablation) | Ours-CJK · Korean (GZSL) | ACR: 19.18 -> 23.50 | Improved |
| double the batch size instead of co-training (ablation) | Ours-CJK · Korean (GZSL) | ACR: 19.18 -> 16.60 | Hurt |
| double the batch size instead of co-training (ablation) | Ours-CJK · Japanese (GZSL) | ACR: 42.36 -> 43.39 | Improved |
| train all Indic scripts as one task rather than separate groups for Punjabi and Tamil (ablation) | Ours-IND · Indic scripts (val) | ACR: Bengali 4.73 -> 0.04; Hindi 82.77 -> 86.35 | Mixed |
| add Indic family co-training to the CJK model (Ours-Uni vs Ours-CJK) | Ours-CJK · Korean and Yi (GZSL) | ACR: Yi 16.10 -> 14.22; Korean 23.32 -> 22.84 | Hurt |
| add CJK co-training to the Indic model (Ours-Uni vs Ours-IND) | Ours-IND · Bengali val (GZSL) | ACR: 8.04 -> 12.88 | Improved |
Closest papers by meaning
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.85)
- Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition (CVPR 2025 · similarity 0.85)
- OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.84)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.84)
- Image as a Language: Revisiting Scene Text Recognition via Balanced, Unified and Synchronized Vision-Language Reasoning Network (AAAI 2024 · similarity 0.83)
- A Simple Framework for Open-Vocabulary Zero-Shot Segmentation (ICLR 2025 · similarity 0.83)
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.
