Research map / Video / Video-language understanding
Does Your Vision-Language Model Get Lost in the Long Video Sampling Dilemma?
Tianyuan Qu, Longxiang Tang, Bohao Peng, Senqiao Yang, Bei Yu, Jiaya Jia
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 SGFS selecting 32 of 128 frames (ablation) | Qwen2.5-VL uniform sampling · LSDBench (Full Video) | accuracy (%): 45.2 -> 48.3 (+3.1) | Improved |
| add SGFS selecting 64 of 128 frames (ablation) | Qwen2.5-VL uniform sampling · LSDBench (Full Video) | accuracy (%): 48.3 -> 47.3 (-1.0) | Hurt |
| add SGFS selecting 64 of 256 frames (ablation) | Qwen2.5-VL uniform sampling · LSDBench (Full Video) | accuracy (%): 48.3 -> 48.7 (+0.4) | No clear effect |
| add SGFS selecting 128 of 512 frames (ablation) | Qwen2.5-VL uniform sampling · LSDBench (Full Video) | accuracy (%): 51 -> 51.8 (+0.8) | Improved |
| add SGFS selecting 256 of 512 frames (ablation) | Qwen2.5-VL uniform sampling · LSDBench (Full Video) | accuracy (%): 50.1 -> 49.6 (-0.5) | Hurt |
| add SGFS selecting 256 of 1024 frames (ablation) | Qwen2.5-VL uniform sampling · LSDBench (Full Video) | accuracy (%): 50.1 -> 51.0 (+0.9) | Improved |
| add SGFS to first-stage sparse sampling (0.25-FPS dense) (ablation) | RHS on Qwen2.5-VL · LSDBench (Full Video, two-stage sparse sampling) | accuracy (%): 49.4 -> 49.9 (+0.5) | Improved |
| add SGFS to first-stage sparse sampling (0.5-FPS dense) (ablation) | RHS on Qwen2.5-VL · LSDBench (Full Video, two-stage sparse sampling) | accuracy (%): 49.4 -> 50.2 (+0.8) | Improved |
| add SGFS to first-stage sparse sampling (1-FPS dense) (ablation) | RHS on Qwen2.5-VL · LSDBench (Full Video, two-stage sparse sampling) | accuracy (%): 51.0 -> 52.2 (+1.2) | Improved |
| replace global uniform 256 frames with RHS (225 frames) | Qwen2.5-VL · LSDBench (Full Video) | accuracy (%): 50.1 -> 52.2 (+2.1) | Improved |
| keep 1.00-FPM instead of 0.50-FPM after SGFS from 4-FPM (ablation) | RHS sparse stage · LSDBench (Full Video, two-stage sparse sampling) | accuracy (%): 51.8 -> 52.2 | Improved |
| raise initial sampling from 2-FPM to 4-FPM (keep 1.00-FPM) (ablation) | RHS sparse stage · LSDBench (Full Video, two-stage sparse sampling) | accuracy (%): 51.5 -> 52.2 | Improved |
Closest papers by meaning
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.91)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.90)
- MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly (NeurIPS 2025 · similarity 0.90)
- SVBench: A Benchmark with Temporal Multi-Turn Dialogues for Streaming Video Understanding (ICLR 2025 · similarity 0.90)
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.89)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · 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.
