Research map / Video / Video-language understanding
Divide, then Ground: Adapting Frame Selection to Query Types for Long-Form Video Understanding
Jialuo Li, Bin Li, Jiahao Li, Yan Lu
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 |
|---|---|---|---|
| use DIG query-adaptive frame selection instead of uniform sampling | Qwen2.5-VL-7B · MLVU (32 frames) | accuracy: 59.52 -> 67.20 | Improved |
| use DIG instead of uniform sampling | Qwen2.5-VL-7B · LongVideoBench (32 frames) | accuracy: 56.92 -> 60.43 | Improved |
| use DIG instead of uniform sampling | Qwen2.5-VL-32B · MLVU (128 frames) | accuracy: 70.24 -> 75.20 | Improved |
| use DIG instead of uniform sampling | Qwen2.5-VL-7B · VideoMME (long, 256 frames) | accuracy: 57.12 -> 57.76 | No clear effect |
| use AKS instead of uniform sampling at high frame counts | Qwen2.5-VL-7B · LongVideoBench (128 frames) | accuracy: 61.86 -> 60.36 | Hurt |
| use Q-Frame instead of uniform sampling | Qwen2.5-VL-32B · MLVU (32 frames) | accuracy: 61.91 -> 60.95 | Hurt |
| increase uniformly sampled input frames beyond an optimum (ablation) | Qwen2.5-VL-7B, InternVL3-8B, LLaVA-OneVision-7B · MLVU, VideoMME, LongVideoBench | accuracy | Hurt |
| replace CLIPScore with LMM (Qwen2.5-VL-7B) reward scoring (ablation) | DIG · MLVU (128 frames) | accuracy: 69.6 -> 71.4 | Improved |
| replace CLIPScore with LMM reward scoring (ablation) | DIG · VideoMME (long, 8 frames) | accuracy: 49.0 -> 46.9 | Hurt |
| use Qwen2.5-VL-32B instead of 7B as reward model (ablation) | DIG · VideoMME (medium, 128 frames) | accuracy: 66.8 -> 69.2 | Improved |
| set window length wlen=0 instead of 2 (ablation) | DIG · MLVU, LongVideoBench, VideoMME | accuracy | Hurt |
| increase window length from 2 to 8 (ablation) | DIG · MLVU, LongVideoBench, VideoMME | accuracy | Hurt |
Closest papers by meaning
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.90)
- Logic-in-Frames: Dynamic Keyframe Search via Visual Semantic-Logical Verification for Long Video Understanding (NeurIPS 2025 · similarity 0.90)
- Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025 · similarity 0.90)
- Adaptive Keyframe Sampling for Long Video Understanding (CVPR 2025 · similarity 0.89)
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.89)
- FlexSelect: Flexible Token Selection for Efficient Long Video Understanding (NeurIPS 2025 · similarity 0.89)
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.
