Research map / Label-efficient and robust learning / Semi-supervised learning
Pushing the Performance Limit of Scene Text Recognizer Without Human Annotation
Caiyuan Zheng, Hui Li, Seon-Min Rhee, Seungju Han, Jae-Joon Han, Peng Wang
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 |
|---|---|---|---|
| Apply consistency regularization using unlabeled real images | CRNN, MORAN, HGA, and TRBA · STR benchmarks (Avg) | accuracy: CRNN: 82.8% to 85.9%; MORAN: 88.5% to 90.2%; HGA: 88.7% to 91.2%; TRBA: 91.5% to 94.5% | Improved |
| Use consistency regularization with 10% synthetic labeled and 10% real unlabeled data | TRBA · STR benchmarks (Avg) | accuracy: 93.2% | Improved |
| Use FixMatch on same-domain labeled and unlabeled data | TRBA · In-domain RL | accuracy: +3.6% | Improved |
| Use UDA on same-domain labeled and unlabeled data | TRBA · In-domain RL | accuracy: +2.6% | Improved |
| Use FixMatch with synthetic labeled and real unlabeled data | TRBA · Cross-domain SLsm and RUsm | accuracy: 11.0% | Hurt |
| Use UDA with synthetic labeled and real unlabeled data | TRBA · Cross-domain SLsm and RUsm | accuracy: 4.6% | Hurt |
| Add a projection layer before the classifier (ablation) | TRBA consistency framework · STR benchmarks (Avg) | accuracy: +0.7% (87.0% to 87.7%) | Improved |
| Add weight decay to the projection-layer variant (ablation) | TRBA consistency framework · STR benchmarks (Avg) | accuracy: +3.5% (87.7% to 91.2%) | Improved |
| Add EMA to the projection-plus-weight-decay variant (ablation) | TRBA consistency framework · STR benchmarks (Avg) | accuracy: +1.6% | Improved |
| Add domain adaptation to the projection-plus-weight-decay-plus-EMA variant (ablation) | TRBA consistency framework · STR benchmarks (Avg) | accuracy: 92.8% to 93.2% | Improved |
| Use character-level consistency regularization instead of standard consistency regularization (ablation) | TRBA · STR benchmarks (Avg) | accuracy: 92.2% to 93.2%; 1% | Improved |
| Replace KL-divergence with MSE for consistency loss (ablation) | TRBA · STR benchmarks (Avg) | accuracy: 93.2% to 91.0% | Hurt |
Closest papers by meaning
- Revisiting Scene Text Recognition: A Data Perspective (ICCV 2023 · similarity 0.91)
- What if We Only Use Real Datasets for Scene Text Recognition? Toward Scene Text Recognition With Fewer Labels (CVPR 2021 · similarity 0.91)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.89)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.88)
- Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition (CVPR 2025 · similarity 0.87)
- TextNeRF: A Novel Scene-Text Image Synthesis Method based on Neural Radiance Fields (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.
