Research map / Large language models / Language, safety and interpretability
Scene Text Recognition with Permuted Autoregressive Sequence Models
Darwin Bautista, Rowel Atienza
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 |
|---|---|---|---|
| train with 6 permutations instead of only left-to-right (K=1) (ablation) | PARSeq-S, NAR decoding · Benchmark (94-char) | word accuracy: 0.01 -> 92.22 | Improved |
| train with 6 permutations instead of only left-to-right (K=1) (ablation) | PARSeq-S, cloze refinement · Benchmark (94-char) | cloze word accuracy: 71.14 -> 94.81 | Improved |
| train with 2 permutations instead of 1 (ablation) | PARSeq-S, AR decoding · Benchmark (94-char) | AR word accuracy: 93.04 -> 93.48 | Improved |
| increase permutations from 6 to 24 (ablation) | PARSeq-S, AR decoding · Benchmark (94-char) | AR word accuracy: 93.34 -> 92.67 | Hurt |
| increase permutations from 6 to 24 (ablation) | PARSeq-S, training cost · Benchmark (94-char) | training hours: 8.48 -> 13.53 | Hurt |
| increase permutations from 6 to 12 (ablation) | PARSeq-S, NAR decoding · Benchmark (94-char) | NAR word accuracy: 92.22 -> 91.71 | Hurt |
| use AR decoding with refinement instead of NAR decoding | PARSeq (synthetic train) · Benchmark mean (94-char) | word accuracy: 70.5 -> 73.0 | Improved |
| train on real data instead of synthetic data | PARSeq_A · Benchmark mean (94-char) | word accuracy: 73.0 -> 93.3 | Improved |
| use PARSeq_A instead of ABINet (external LM) with synthetic training | ABINet baseline · Benchmark mean (94-char) | word accuracy: 66.4 -> 73.0 | Improved |
| use PARSeq_A instead of ABINet with real training on harder datasets | ABINet baseline · ArT+COCO+Uber total (36-char) | word accuracy: 74.6 -> 84.1 | Improved |
| add flipped (reversed) permutation pairs to sampling (ablation) | PARSeq training · Training stability | training loss smoothness | Improved |
| add position-query tokens decoupled from context (ablation) | PARSeq decoder · PLM training | learning from PLM | Improved |
Closest papers by meaning
- Revisiting Scene Text Recognition: A Data Perspective (ICCV 2023 · similarity 0.90)
- What if We Only Use Real Datasets for Scene Text Recognition? Toward Scene Text Recognition With Fewer Labels (CVPR 2021 · similarity 0.89)
- What Is Wrong With Scene Text Recognition Model Comparisons? Dataset and Model Analysis (ICCV 2019 · similarity 0.88)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.87)
- Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition (CVPR 2025 · similarity 0.86)
- 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.
