Research map / Video / Video-language understanding
STVFocus: Query-guided Spatio-Temporal Visual Focusing for Video LLMs
Taehun Kong, Minyoung Park, Sangjun Ahn
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 coarse (diverse) frame selection (ablation) | LLaVA-OneVision-7B · VideoMME | accuracy: 58.5 -> 59.2 | Improved |
| add fine (query-relevant) frame selection (ablation) | LLaVA-OneVision-7B · VideoMME | accuracy: 58.5 -> 60.2 | Improved |
| combine coarse and fine selection (coarse-to-fine) (ablation) | LLaVA-OneVision-7B · VideoMME | accuracy: 58.5 -> 60.8 | Improved |
| add QSF with 80% token budget on top of HTF (ablation) | LLaVA-OneVision-7B · VideoMME | accuracy: 60.8 -> 60.8 | No clear effect |
| add temporal reallocation (32+8 frames) (ablation) | LLaVA-OneVision-7B · VideoMME | accuracy: 60.8 -> 61.1 | Improved |
| use AAM only for spatial focusing at 50% budget (ablation) | LLaVA-OneVision-7B · VideoMME | accuracy: 58.5 -> 59.1 | Improved |
| use spatial saliency only at 50% budget (ablation) | LLaVA-OneVision-7B · VideoMME | accuracy: 58.5 -> 58.4 | No clear effect |
| fuse AAM with spatial saliency (ablation) | LLaVA-OneVision-7B · VideoMME | accuracy: 59.1 -> 59.4 | Improved |
| set gamma 0.5 instead of 0.7 in QSF (ablation) | QSF · VideoMME | accuracy: 58.6 -> 59.4 | Improved |
| set tau 80% instead of 90% in QSF (ablation) | STVFocus · VideoMME | accuracy: 60.5 -> 61.1 | Improved |
| set beta 0.5 instead of 0.3 in HTF fine phase (ablation) | STVFocus · VideoMME | accuracy: 59.0 -> 60.2 | Improved |
| add STVFocus under 8 frames | InternVL3-8B · VideoMME (8 frames) | accuracy: 59.8 -> 63.3 | Improved |
Closest papers by meaning
- SlowFocus: Enhancing Fine-grained Temporal Understanding in Video LLM (NeurIPS 2024 · similarity 0.89)
- Unleashing the Potential of Multimodal LLMs for Zero-Shot Spatio-Temporal Video Grounding (NeurIPS 2025 · similarity 0.88)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.88)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.87)
- SF2T: Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding (CVPR 2025 · similarity 0.87)
- DELTAVID: Enhancing Fine-Grained Spatiotemporal Perception with Cross-Video Differences (NeurIPS 2026 · similarity 0.87)
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.
