Research map / Vision-language and multimodal / Human and embodied modeling
Scene Text Retrieval via Joint Text Detection and Similarity Learning
Hao Wang, Xiang Bai, Mingkun Yang, Shenggao Zhu, Jing Wang, Wenyu Liu
Check 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 |
|---|---|---|---|
| Jointly optimize text detection and cross-modal similarity learning instead of separating the stages (ablation) | Scene text retrieval · SVT, STR, CTR | mAP: 82.88 -> 89.38; 70.83 -> 77.09; 61.16 -> 66.45 | Improved |
| Add word augmentation strategy (ablation) | Baseline similarity-learning method · SVT, STR, CTR | mAP: 88.05 -> 88.68; 73.70 -> 75.76; 62.97 -> 64.59 | Improved |
| Add CTC loss to align visual and textual features (ablation) | Baseline similarity-learning method · SVT, STR, CTR | mAP: 88.05 -> 88.62; 73.70 -> 75.54; 62.97 -> 64.08 | Improved |
| Add both word augmentation strategy and CTC loss (ablation) | Baseline similarity-learning method · SVT, STR, CTR | mAP: 88.05 -> 89.38; 73.70 -> 77.09; 62.97 -> 66.45 | Improved |
| Add auxiliary similarity training on proposal-proposal and query-query pairs (ablation) | Scene text retrieval model · SVT, STR, CTR | mAP: 85.28 -> 87.78; 66.50 -> 70.18; 51.79 -> 53.90 | Improved |
| Learn cross-modal similarity instead of predicting PHOC vectors (ablation) | Same text detection network · SVT, STR, CTR | mAP: 80.87 -> 87.78; 60.00 -> 70.18; 42.89 -> 53.90 | Improved |
| Use multi-scale testing | Ours · SVT, STR, CTR | mAP: +2.31%; +3.07%; +3.42% | Improved |
| Use the proposed retrieval method rather than the strongest prior method in the comparison | Scene text retrieval · SVT, STR, CTR | mAP: 85.74 -> 89.38; 74.48 -> 77.09; 55.54 -> 66.45 | Improved |
| Apply the method to Chinese retrieval with 3755 character categories rather than 1019 | Ours on CSVTR · CSVTR | mAP: 60.23 -> 50.12 | Hurt |
Closest papers by meaning
- Unambiguous Text Localization, Retrieval, and Recognition for Cluttered Scenes (TPAMI 2020 · similarity 0.88)
- Unambiguous Text Localization and Retrieval for Cluttered Scenes (CVPR 2017 · similarity 0.88)
- MSTAR: Box-free Multi-query Scene Text Retrieval with Attention Recycling (NeurIPS 2025 · similarity 0.88)
- ViSTA: Vision and Scene Text Aggregation for Cross-Modal Retrieval (CVPR 2022 · similarity 0.88)
- Heterogeneous Prompt-Guided Entity Inferring and Distilling for Scene-Text Aware Cross-Modal Retrieval (AAAI 2025 · similarity 0.86)
- CLIP is Almost All You Need: Towards Parameter-Efficient Scene Text Retrieval without OCR (CVPR 2025 · similarity 0.86)
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.
