Research map / Architectures and efficiency / RNNs and normalization
An Empirical Study of Scaling Law for Scene Text Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Empirical study showing smooth power laws between scene text recognition performance and model size, data volume, and compute for TrOCR and PARSeq, and builds the REBU-Syn dataset to train CLIP4STR-L to 97.42% average word accuracy on 6 benchmarks.
Problem
Scaling laws relating model size, training data volume, and computation to performance have been studied in NLP but not in scene text recognition (STR).
Approach
Scales transformer-based STR models (TrOCR and PARSeq) across sizes and data volumes, fits power laws to word error rate, and combines public real and synthetic data (REBU-Syn, 3:1 synthetic-to-real ratio) with STR-task pretraining to train CLIP4STR-L to a new state of the art.
Key results
- CLIP4STR-L trained on REBU-Syn reaches 97.42% weighted average word accuracy over 6 common STR benchmarks, a new state of the art in the paper's comparison.
- Fitted power laws: TrOCR error vs. parameters E(N)=(1.97e4/N)^0.223, and vs. data volume E(D)=(1.84e5/D)^-0.327.
- PARSeq-S trained on REBU-Syn reaches 96.85% average accuracy, versus 96.12% when trained from scratch on real data only (R+E+B+U) with ViT-S.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train CLIP4STR-B with REBU-Syn instead of real data | CLIP4STR-B · 6 common test benchmarks | average word accuracy: 96.54% to 97.25% (+0.65%) | Helped |
| Train CLIP4STR-L with REBU-Syn | CLIP4STR-L · 6 common test benchmarks | average word accuracy: 97.06 to 97.42 | Helped |
| Increase TrOCR model size from TrOCR-S to TrOCR-H | TrOCR · 6 common test benchmarks | average word accuracy: 81.93 -> 90.94 | Helped |
| Increase PARSeq model size from PARSeq-S to PARSeq-H | PARSeq · 6 common test benchmarks | average word accuracy: 96.85 -> 97.06 | Helped |
| Increase the synthetic-to-real data ratio from 1:3 to 1:5 (ablation) | PARSeq-S · 6 common test benchmarks | average word accuracy: 96.85 -> 95.70 | Hurt |
| Add synthetic MJ+ST data to real REBU data (ablation) | PARSeq-S · 6 common test benchmarks | average word accuracy: 95.57 -> 96.24 | Helped |
| Add Syn data to real REBU data (ablation) | PARSeq-S · 6 common test benchmarks | average word accuracy: 96.19 | Helped |
| Combine real data with both MJ+ST and Syn synthetic data (ablation) | PARSeq-S · 6 common test benchmarks | average word accuracy: 96.24 -> 96.85 | Helped |
| Pretrain on ImageNet-21k rather than train from scratch (ablation) | PARSeq-L · 6 common test benchmarks | average word accuracy: 97.03 -> 96.74 | Hurt |
| Pretrain on STR-related data and fine-tune on real data (ablation) | PARSeq-S · 6 common test benchmarks | average word accuracy: 96.85% to 97.01% | Helped |
Surprising findings
- ImageNet-21k pretrained ViT-L underperformed training from scratch for PARSeq (96.74% vs 97.03%), suggesting general visual pretraining may hurt STR.
- Synthetic data beyond a 1:3 synthetic-to-real ratio reduced accuracy (95.70% at 1:5), indicating skew toward synthetic data is harmful.
Limitations
- Experiments limited to natural scene text; handwriting and historical text not studied.
- Power-law fits use few model sizes (TrOCR-H fitted as validation) and large-model training is very costly.
- Scaling-law fits are reported for specific architectures (TrOCR, PARSeq) and may not transfer to other STR models.
Experimental setup
Datasets: REBU-Syn Union14M-L COCO-Text RCTW17 Uber-Text ArT LSVT MLT19 ReCTS TextOCR OpenVINO IIIT5k
Benchmarks: IIIT5k SVT IC13 IC15 SVTP CUTE80
Metrics: word accuracy average word accuracy average word error rate
Baselines: TrOCR PARSeq CLIP4STR-B CLIP4STR-L MaskOCR MAERec ABINet SRN DiG-ViT-B
Builds on: TrOCR PARSeq CLIP4STR Union14M
Closest papers by meaning
- Scaling Laws for Robust Comparison of Open Foundation Language-Vision Models and Datasets (NeurIPS 2025 · similarity 0.88)
- Accurate Scene Text Recognition with Efficient Model Scaling and Cloze Self-Distillation (CVPR 2025 · similarity 0.87)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.86)
- Scaling Laws for Optimal Data Mixtures (NeurIPS 2025 · similarity 0.86)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.85)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.85)
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.