Research map / Vision-language and multimodal / CLIP and zero-shot
Towards the Unseen: Iterative Text Recognition by Distilling From Errors
Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Yi-Zhe Song
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 CVAE feedback (t=2) | Shi et al. · IAM (disjoint setup) | word recognition accuracy: 54.3 -> 64.8 | Improved |
| add iterative CVAE feedback (t=2) | Shi et al. · RIMES (disjoint setup) | word recognition accuracy: 59.7 -> 70.5 | Improved |
| add iterative CVAE feedback (t=2) | Shi et al. · IC15 (conventional setup) | word recognition accuracy: 75.9 -> 78.8 | Improved |
| add iterative CVAE feedback (t=2) | Show, Attend and Read · IC15 (disjoint setup) | word recognition accuracy: 68.4 -> 74.8 | Improved |
| add iterative CVAE feedback (t=2) | SCATTER · IC15 (disjoint setup) | word recognition accuracy: 71.8 -> 77.7 | Improved |
| run a third feedback iteration (t=3 instead of t=2) (ablation) | Shi et al. + CVAE-Feedback · IC15 (disjoint setup) | word recognition accuracy: 72.6 -> 72.5 | No clear effect |
| use CVAE feedback instead of deterministic feedback | Shi et al. · IC15 (disjoint setup) | word recognition accuracy: 70.4 -> 72.6 | Improved |
| add auxiliary decoder and error-distribution data to feedback (vs L_C only) (ablation) | CVAE-Feedback · IAM (disjoint setup) | word recognition accuracy: 59.8 -> 64.8 | Improved |
| add monotonically decreasing constraint L_C (ablation) | CVAE-Feedback · IAM (disjoint setup) | word recognition accuracy: 59.3 -> 64.8 | Improved |
| add error distribution data G (ablation) | CVAE-Feedback · IAM (disjoint setup) | word recognition accuracy: 59.9 -> 64.8 | Improved |
| use local instead of global feedback transformation (block 3) (ablation) | CVAE-Feedback · IC15 (disjoint setup) | word recognition accuracy: 71.4 -> 72.6 | Improved |
| apply feedback at ResNet block 3 rather than block 1 (ablation) | CVAE-Feedback · IAM (disjoint setup) | word recognition accuracy: 61.0 -> 64.8 | Improved |
Closest papers by meaning
- OCRBench v2: An Improved Benchmark for Evaluating Large Multimodal Models on Visual Text Localization and Reasoning (NeurIPS 2025 · similarity 0.88)
- ASCIIEval: Benchmarking Models' Visual Perception in Text Strings via ASCII Art (ICLR 2026 · similarity 0.88)
- T-MARS: Improving Visual Representations by Circumventing Text Feature Learning (ICLR 2024 · 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 0.87)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.87)
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.
