Research map / Vision-language and multimodal / Human and embodied modeling
Towards Unified Scene Text Spotting Based on Sequence Generation
Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, Daehee Kim
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 starting-point prompting at test time (iterative re-prompting from last detected text) (ablation) | UNITS on TextOCR validation · TextOCR validation | End-to-end F-measure: 44.0 -> 64.7 | Improved |
| add starting-point prompting (ablation) | UNITS on TextOCR validation · TextOCR validation | End-to-end Recall: 30.6 -> 54.2 | Improved |
| reduce decoder length from 1024 to 256 with starting-point prompting (ablation) | UNITS on TextOCR validation · TextOCR validation | End-to-end F-measure: 64.7 -> 61.8 | Mixed |
| replace Swin encoder with hybrid ResNet-50 + transformer encoder (ablation) | UNITS · ICDAR 2015 and Total-Text | End-to-end E2E: 78.5 -> 65.6 (ICDAR 2015); 77.3 -> 67.4 (Total-Text) | Hurt |
| add multi-way (detection/recognition expert FFN) decoder (ablation) | UNITS decoder · ICDAR 2015 | End-to-end E2E: 39.1 -> 40.7 | Improved |
| add multi-way (detection/recognition expert FFN) decoder (ablation) | UNITS decoder · Total-Text | End-to-end E2E: 45.1 -> 47.7 | Improved |
| add multi-way decoder (ablation) | UNITS decoder · ICDAR 2015 | Detection: 73.7 -> 75.1 | Improved |
| add multi-way decoder (ablation) | UNITS decoder · Total-Text | Detection: 76.1 -> 78.9 | Improved |
| unified multi-format training with a single shared model instead of per-dataset fine-tuned models | UNITS · ICDAR 2015 End-to-end Strong | End-to-end Strong F-measure: 89.0 -> 88.4 | Hurt |
| unified multi-format model vs dataset-specific fine-tuning on Total-Text | UNITS · Total-Text | End-to-end Full F-measure: 86.0 -> 85.0 | Hurt |
| use the sequence generation model with starting-point prompting and unified multi-format interface, compared to SPTS on ICDAR 2015 point metric | UNITS_Shared (quad) vs SPTS · ICDAR 2015 (point-based metric) | End-to-end Strong: 77.5 -> 89.9 | Improved |
Closest papers by meaning
- SPTS v2: Single-Point Scene Text Spotting (TPAMI 2023 · similarity 0.89)
- Mask TextSpotter v3: Segmentation Proposal Network for Robust Scene Text Spotting (ECCV 2020 · similarity 0.88)
- All You Need Is Boundary: Toward Arbitrary-Shaped Text Spotting (AAAI 2020 · similarity 0.88)
- Character Region Attention For Text Spotting (ECCV 2020 · similarity 0.88)
- SemiETS: Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting (CVPR 2025 · similarity 0.86)
- Arbitrary Reading Order Scene Text Spotter with Local Semantics Guidance (AAAI 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.
