Research map / Vision-language and multimodal / Human and embodied modeling
Levenshtein OCR
Cheng Da, Peng Wang, 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 |
|---|---|---|---|
| add iterative refinement from visual-prediction initial sequence (3 iterations) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST) | word accuracy: 90.65 -> 92.81 | Improved |
| start from empty initial sequence (3 iterations) (ablation) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST) | word accuracy: 90.65 -> 92.63 | Improved |
| use visual prediction rather than empty initial sequence (ablation) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST) | word accuracy: 92.63 -> 92.81 | Improved |
| use random-noise corrupted sequence instead of visual prediction as y0 (ablation) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST) | word accuracy: 92.81 -> 85.25 | Hurt |
| use ground-truth initial sequence (ablation) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST) | word accuracy: 92.81 -> 96.91 | Improved |
| second refinement iteration with visual-prediction init (ablation) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST) | word accuracy: 92.55 -> 92.79 | Improved |
| second iteration with empty initial sequence (ablation) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST) | word accuracy: 90.64 -> 92.51 | Improved |
| third iteration over second (visual-prediction init) (ablation) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST) | word accuracy: 92.79 -> 92.81 | No clear effect |
| add Levenshtein Transformer deletion and insertion losses (ablation) | LevOCR (CNN) · average of 6 benchmarks (MJ+ST), backbone study | word accuracy: 90.70 -> 92.81 | Improved |
| add LevT refinement with ViT backbone | LevOCR (ViT) · average of 6 benchmarks (MJ+ST), backbone study | word accuracy: 89.30 -> 91.87 | Improved |
| use LevOCR instead of SRN | scene text recognizer · average of 6 benchmarks | word accuracy: 90.4 -> 92.81 | Improved |
| use LevOCR instead of ABINet | scene text recognizer · average of 6 benchmarks | word accuracy: 92.6 -> 92.81 | Mixed |
Closest papers by meaning
- From Two to One: A New Scene Text Recognizer With Visual Language Modeling Network (ICCV 2021 · similarity 0.86)
- Language Matters: A Weakly Supervised Vision-Language Pre-training Approach for Scene Text Detection and Spotting (ECCV 2022 · similarity 0.86)
- TextOCR: Towards Large-Scale End-to-End Reasoning for Arbitrary-Shaped Scene Text (CVPR 2021 · similarity 0.86)
- Multi-Granularity Prediction for Scene Text Recognition (ECCV 2022 · similarity 0.85)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.85)
- Image as a Language: Revisiting Scene Text Recognition via Balanced, Unified and Synchronized Vision-Language Reasoning Network (AAAI 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.
