Research map / Detection and segmentation / CNN features
Dynamic Low-Resolution Distillation for Cost-Efficient End-to-End Text Spotting
Ying Chen, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, Xi Li
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 sequential knowledge distillation at half resolution (ablation) | Vanilla Multi-Scale low-res · Total-Text | End-to-End None: 55.4 -> 59.6 | Improved |
| add sequence-level beam search loss L_seq alone (ablation) | SKD-only · Total-Text | End-to-End None: 55.4 -> 58.4 | Improved |
| add all three SKD losses (ablation) | SKD-only · Total-Text | End-to-End None: 55.4 -> 59.6 | Improved |
| replace sequence-level distillation with logits-based distillation (ablation) | SKD · Total-Text | End-to-End None: 59.6 -> 58.7 | Hurt |
| add detection distillation of Qi et al. to SKD (ablation) | SKD · Total-Text | End-to-End None: 59.6 -> 59.8 | No clear effect |
| add dynamic resolution selector to SKD (full DLD, gamma 0.1) | SKD-only · Total-Text | End-to-End None: 59.6 -> 63.9 | Improved |
| use DLD instead of high-res vanilla input | Vanilla Multi-Scale · IC15 | End-to-End General: 69.5 -> 70.9 | Improved |
| use larger RoI scale 16x64 in teacher and student (ablation) | DLD · Total-Text | End-to-End None: 63.9 -> 64.8 | Improved |
| align smaller student RoI with deconv instead of using larger RoI (ablation) | DLD · Total-Text | End-to-End None: 64.8 -> 64.5 | Hurt |
| use larger candidate scales {0.8,0.7,0.6} instead of {0.5,0.4,0.3} (ablation) | DLD · Total-Text | End-to-End None: 58.9 -> 63.6 | Improved |
| use six candidate scales instead of three larger ones (ablation) | DLD · Total-Text | End-to-End None: 63.6 -> 63.9 | Improved |
| increase gamma from 0.1 to 0.5 (ablation) | DLD · Total-Text | End-to-End None: 63.9 -> 56.0 | Hurt |
Closest papers by meaning
- Bridging the Gap Between End-to-End and Two-Step Text Spotting (CVPR 2024 · similarity 0.88)
- Text Is Text, No Matter What: Unifying Text Recognition Using Knowledge Distillation (ICCV 2021 · similarity 0.87)
- MANGO: A Mask Attention Guided One-Stage Scene Text Spotter (AAAI 2021 · similarity 0.86)
- GoMatching: A Simple Baseline for Video Text Spotting via Long and Short Term Matching (NeurIPS 2024 · similarity 0.85)
- ODM: A Text-Image Further Alignment Pre-training Approach for Scene Text Detection and Spotting (CVPR 2024 · similarity 0.85)
- SemiETS: Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting (CVPR 2025 · similarity 0.84)
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.
