Research map / Vision-language and multimodal / Human and embodied modeling
LISTER: Neighbor Decoding for Length-Insensitive Scene Text Recognition
Changxu Cheng, Peng Wang, Cheng Da, Qi Zheng, Cong Yao
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 |
|---|---|---|---|
| replace the neighbor decoder with a parallel attention decoder (ablation) | LISTER-T* · TUL | accuracy (%): 76.8 -> 61.6 | Hurt |
| use neighbor decoder instead of serial attention decoder (ablation) | LISTER-T* · TUL | accuracy (%): 66.5 -> 76.8 | Improved |
| use neighbor decoder instead of CTC (ablation) | LISTER-T* · common benchmarks | average accuracy (%): 90.2 -> 91.7 | Improved |
| add attention sharpening at inference (ablation) | Neighbor Decoder · TUL unseen lengths (trained on lengths <=16) | accuracy (%): 47.1 -> 58.4 | Improved |
| add Feature Enhancement Module on top of attention sharpening (ablation) | Neighbor Decoder · TUL unseen lengths (trained on lengths <=16) | accuracy (%): 58.4 -> 61.4 | Improved |
| add both Transformer layer and convolution block in FEM (ablation) | LISTER-T · common benchmarks | average accuracy (%): 92.0 / 92.2 -> 92.8 | Improved |
| increase FEM iterations from 1 to 2 (2 Transformer, 2 conv) (ablation) | LISTER-T · TUL | accuracy (%): 77.2 -> 77.8 | Improved |
| place Transformer layers at the end of the feature extractor instead of using FEM (ablation) | LISTER-T · common benchmarks | average accuracy (%): 92.8 -> 92.0 | Hurt |
| pad images instead of resizing to fixed width (ablation) | Neighbor Decoder · TUL | accuracy (%): 65.9 -> 76.8 | Improved |
| scale model from LISTER-T* to LISTER-B | LISTER · TUL | accuracy (%): 76.8 -> 79.2 | Improved |
| use a feature map of height 4 instead of 1 (ablation) | LISTER-B · ArT (synthetic training) | average accuracy (%): 56.2 -> 57.4 | Improved |
| add 3-scale test-time ensemble | LISTER-B · common benchmarks | average accuracy (%): 93.3 -> 93.5 | Improved |
Closest papers by meaning
- Out of Length Text Recognition with Sub-String Matching (AAAI 2025 · similarity 0.88)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.88)
- RobustScanner: Dynamically Enhancing Positional Clues for Robust Text Recognition (ECCV 2020 · similarity 0.88)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.86)
- CLIP is Almost All You Need: Towards Parameter-Efficient Scene Text Retrieval without OCR (CVPR 2025 · similarity 0.86)
- Accurate Scene Text Recognition with Efficient Model Scaling and Cloze Self-Distillation (CVPR 2025 · 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.
