Research map / Architectures and efficiency / Attention and transformers
SCATTER: Selective Context Attentional Scene Text Recognizer
Ron Litman, Oron Anschel, Shahar Tsiper, Roee Litman, Shai Mazor, R. Manmatha
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 SCATTER with 5 blocks rather than Mask TextSpotter | ICDAR2015 recognition · IC15 | accuracy (%): 78.2% vs. 82.2%; +4.0 pp | Improved |
| Use SCATTER with 5 blocks rather than Mask TextSpotter | SVTP recognition · SVTP | accuracy (%): 83.6% vs. 86.9%; +3.3 pp | Improved |
| Use SCATTER with 5 blocks rather than Mask TextSpotter | CUTE recognition · CUTE | accuracy (%): 88.5% vs. 87.5% | Hurt |
| Add intermediate CTC supervision (ablation) | SCATTER 1-block model · Regular text weighted average | accuracy (%): +0.2 pp | Improved |
| Add intermediate CTC supervision (ablation) | SCATTER 1-block model · Irregular text weighted average | accuracy (%): +0.4 pp | Improved |
| Replace the standard attentional decoder with the selective decoder (ablation) | SCATTER 1-block model · Regular text weighted average | accuracy (%): +0.4 pp | Improved |
| Replace the standard attentional decoder with the selective decoder (ablation) | SCATTER 1-block model · Irregular text weighted average | accuracy (%): +2.7 pp | Improved |
| Increase the number of intermediate selective decoders from 1 to 3 (ablation) | SCATTER with 4 BiLSTM layers · Regular text weighted average | accuracy (%): +0.7 pp | Improved |
| Increase the number of intermediate selective decoders from 1 to 3 (ablation) | SCATTER with 4 BiLSTM layers · Irregular text weighted average | accuracy (%): +2.9 pp | Improved |
| Add intermediate supervision and stack more BiLSTM layers (ablation) | SCATTER encoder · Regular and irregular text weighted averages | accuracy (%): 92.7 -> 94.0; 79.1 -> 83.7 | Improved |
| Increase encoder depth beyond 10 BiLSTM layers (ablation) | SCATTER · Regular and irregular text weighted averages | accuracy (%): -0.4 pp; -0.5 pp | Hurt |
| Train with 5 blocks, then prune to one block for inference (ablation) | SCATTER · Regular text weighted average | accuracy (%): +0.4 pp | Improved |
Closest papers by meaning
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.86)
- Instruction-Guided Scene Text Recognition (TPAMI 2025 · similarity 0.86)
- Background-Insensitive Scene Text Recognition with Text Semantic Segmentation (ECCV 2022 · similarity 0.86)
- Context-Based Contrastive Learning for Scene Text Recognition (AAAI 2022 · similarity 0.86)
- Accurate Scene Text Recognition with Efficient Model Scaling and Cloze Self-Distillation (CVPR 2025 · similarity 0.85)
- CLIP is Almost All You Need: Towards Parameter-Efficient Scene Text Retrieval without OCR (CVPR 2025 · similarity 0.84)
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.
