Research map / Vision-language and multimodal / Human and embodied modeling
Mind the Gap: Quantifying and Aligning Human-AI Visual Attention for Accident Anticipation
Hoe Sung Ryu, Christian Wallraven
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 prompt-based attention guidance (ablation) | VLMs · Attention-DADA | macro F1: open-source about +7%, closed-source +6.5% | Improved |
| add visual-based human gaze guidance (ablation) | VLMs · Attention-DADA | macro F1: open-source about +4%, closed-source <1% | Improved |
| add prompt guidance to Gemini-2.5-Flash (ablation) | Gemini-2.5-Flash · Attention-DADA T1 | macro F1: 0.62 -> 0.67 | Improved |
| add prompt guidance to QwenVL-2.5 (ablation) | QwenVL-2.5 · Attention-DADA T1 | macro F1: 0.42 -> 0.43 | No clear effect |
| add attention guidance (prompt or visual) (ablation) | VLMs · Attention-DADA | gap to human F1 near event: human 0.88 vs best VLM 0.67 at T1 | No clear effect |
| add attention guidance (prompt or visual) (ablation) | InternVL-3 · Attention-DADA | Kendall's tau alignment | No clear effect |
| move from earliest (T5) to latest (T1) clip | human participants · Attention-DADA | macro F1: 0.46 -> 0.88 | Improved |
| move from earliest (T5) to latest (T1) clip | zero-shot VLMs · Attention-DADA | macro F1: e.g. QwenVL-2.5 0.38 -> 0.42 | No clear effect |
| accuracy level (high vs low) for prompt-guided model on normal videos | prompt-guided VLM · Attention-DADA normal videos | Kendall's tau: 0.355 -> 0.481 | Improved |
| accuracy level vs tau on other five conditions | VLMs · Attention-DADA | Kendall's tau | No clear effect |
Closest papers by meaning
- PSI: A Benchmark for Human Interpretation and Response in Traffic Interactions (NeurIPS 2025 · similarity 0.85)
- Unifying Top-down and Bottom-up Scanpath Prediction Using Transformers (CVPR 2024 · similarity 0.84)
- Whose View of Safety? A Deep DIVE Dataset for Pluralistic Alignment of Text-to-Image Models (NeurIPS 2025 · similarity 0.84)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.84)
- Learning from Observer Gaze: Zero-Shot Attention Prediction Oriented by Human-Object Interaction Recognition (CVPR 2024 · similarity 0.84)
- Accident Anticipation via Temporal Occurrence Prediction (NeurIPS 2025 · similarity 0.83)
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.
