SciFigQual-Bench: Benchmark for Scientific Figure Quality Assessment
Researchers propose SciFigQual-Bench, a benchmark for assessing scientific image quality using full-manuscript context. It evaluates images across five dimensions: clarity, layout, caption fit, context relevance, and misleading risk. The dataset includes 6,308 images from top computer science conferences (2020–2025), scored by multiple experts. This addresses the gap where existing IQA methods fail for scientific papers.
Key facts
- SciFigQual-Bench evaluates scientific images across five dimensions: clarity, layout, caption fit, context relevance, and misleading risk.
- The dataset covers top computer-science conferences from 2020 to 2025.
- 6,308 images were independently scored by multiple domain experts.
- Existing IQA methods are designed for natural photographs or AI-generated content.
- Previous studies on scholarly charts focus on visual-surface comparisons only.
- The benchmark uses full-text context to verify caption alignment, citation relevance, and visual misleadingness.
- Gold-standard annotations were aggregated from expert scores.
- The paper is available on arXiv (2607.27084).
Entities
Institutions
- arXiv