Research map / Architectures and efficiency / CNN design
Pure Transformer with Integrated Experts for Scene Text Recognition
Yew Lee Tan, Adams Wai-Kin Kong, Jung-Jae Kim
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 a pure ViT encoder instead of ResNet-45 plus transformer encoder (ablation) | transformer STR model · 6 benchmarks weighted avg (7672) | accuracy: 85.7 -> 90.9 | Improved |
| train experts with diverse patch resolutions and ground-truth orders instead of identical ones (ablation) | PTIE · 6 benchmarks weighted avg (7672) | accuracy: 91.0 -> 92.4 | Improved |
| integrate four experts in one model (PTIE-Diverse) instead of one 8x4 model | ViT baseline · 6 benchmarks weighted avg (7672) | accuracy: 90.9 -> 92.4 | Improved |
| replace four-model ensemble with PTIE-Diverse | Ensemble-Diverse · 6 benchmarks weighted avg (7672) | accuracy: 92.4 -> 92.4 | No clear effect |
| replace four-model ensemble with PTIE-Diverse | Ensemble-Diverse · latency | inference time (ms): 202 -> 52 | Improved |
| use diverse instead of identical members in the ensemble (ablation) | Ensemble · 6 benchmarks weighted avg (7672) | accuracy: 92.1 -> 92.4 | Improved |
| add PTIE integrated experts to the baseline | ViT baseline · latency | inference time (ms): 50 -> 52 | No clear effect |
| use untied positional attention instead of vanilla positional encoding (ablation) | PTIE · IC15 (2077) | accuracy: 83.4 -> 84.3 | Improved |
| use untied positional attention instead of vanilla positional encoding (ablation) | PTIE · IIIT5K | accuracy: 96.7 -> 96.3 | Hurt |
| use untied positional attention instead of vanilla positional encoding (ablation) | PTIE · weighted average over benchmarks | accuracy: 0.1% higher | Mixed |
| fine-tune with real data COCO-Text | PTIE-Untied · SVT-P | accuracy: 90.1 -> 92.1 | Improved |
| fine-tune with real data COCO-Text | PTIE-Untied · CUTE80 | accuracy: 91.7 -> 91.0 | Hurt |
Closest papers by meaning
- Multi-Granularity Prediction for Scene Text Recognition (ECCV 2022 · similarity 0.88)
- Multi-Granularity Prediction with Learnable Fusion for Scene Text Recognition (IJCV 2026 · similarity 0.87)
- ViTEraser: Harnessing the Power of Vision Transformers for Scene Text Removal with SegMIM Pretraining (AAAI 2024 · similarity 0.87)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.87)
- What Is Wrong With Scene Text Recognition Model Comparisons? Dataset and Model Analysis (ICCV 2019 · similarity 0.86)
- 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.
