Research map / Vision-language and multimodal / Human and embodied modeling
Attention Where It Matters: Rethinking Visual Document Understanding with Selective Region Concentration
Haoyu Cao, Changcun Bao, Chaohu Liu, Huang Chen, Kun Yin, Hao Liu, Yinsong Liu, Deqiang Jiang, Xing Sun
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 text-to-segmentation pre-training task to query-to-segmentation (ablation) | SeRum · SROIE | F1: 59.3 -> 82.5 | Improved |
| add segmentation-to-text pre-training task (ablation) | SeRum · SROIE | F1: 82.5 -> 85.8 | Improved |
| use prompt-based parallel generation (SeRum-prompt) instead of one-pass total generation (ablation) | SeRum · CORD | F1: 80.5 -> 84.9 | Improved |
| use prompt-based generation instead of total generation (ablation) | SeRum · Ticket | F1: 97.9 -> 99.2 | Improved |
| add content-aware token merge (keep ratio 10%) versus no merge (ablation) | SeRum · SROIE | F1: 84.9 -> 85.8 | Improved |
| add content-aware token merge (keep ratio 10%) versus no merge (ablation) | SeRum · SROIE | text decoder latency (ms): 306 -> 209 | Improved |
| reduce token keep ratio too far (2%) (ablation) | SeRum · SROIE | F1: 85.8 -> 72.5 | Hurt |
| end-to-end SeRum versus OCR-based LayoutLMv2 | document extraction · SROIE | F1: 61.0 -> 85.8 | Improved |
| SeRum on text spotting versus specialist SPTS v2 | SeRum · CTW-1500 text spotting | F1: 63.6 -> 41.8 | Hurt |
| SeRum versus OCR-based LayoutLMv2 on DocVQA ANLS | SeRum · DocVQA | ANLS: 78.1 vs 71.9 | Hurt |
Closest papers by meaning
- DocKylin: A Large Multimodal Model for Visual Document Understanding with Efficient Visual Slimming (AAAI 2025 · similarity 0.86)
- Hierarchical Visual Feature Aggregation for OCR-Free Document Understanding (NeurIPS 2024 · similarity 0.85)
- HRVDA: High-Resolution Visual Document Assistant (CVPR 2024 · similarity 0.84)
- Boosting Document Parsing Efficiency and Performance with Coarse-to-Fine Visual Processing (CVPR 2026 · similarity 0.84)
- DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding (CVPR 2025 · similarity 0.84)
- Focus, Don't Prune: Identifying Instruction-Relevant Regions for Information-Rich Image Understanding (CVPR 2026 · 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.
