Research map / Video / Video-language understanding
APVR: Hour-Level Long Video Understanding with Adaptive Pivot Visual Information Retrieval
Hong Gao, Yiming Bao, Xuezhen Tu, Bin Zhong, Linan Yue, Min-Ling Zhang
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 expanded semantic information to the query (ablation) | APVR Qwen2-VL-7B · LongVideoBench | Avg accuracy: 59.1 -> 60.9 | Improved |
| add expanded semantic information to the query (ablation) | APVR Qwen2.5-VL-7B · LongVideoBench | Avg accuracy: 63.2 -> 64.5 | Improved |
| add expanded semantic information to the query (ablation) | APVR VideoLLaMA3-7B · LongVideoBench | Avg accuracy: 62.3 -> 63.5 | Improved |
| increase pivot frames K from 256 to 1024 (ablation) | APVR · LongVideoBench | Avg accuracy: 63.1 -> 64.9 | Improved |
| increase pivot frames K from 32 to 1024 (ablation) | APVR · LongVideoBench | Avg accuracy: 54.3 -> 64.9 | Improved |
| balance CLIP and Grounding-DINO weight (lambda 0.5 vs 0.1) (ablation) | APVR · LongVideoBench | Avg accuracy: 64.2 -> 64.9 | Improved |
| use 3 iterations with stride 4 instead of 2 iterations (ablation) | APVR · LongVideoBench | Avg accuracy: 63.9 -> 64.9 | Improved |
| use 5 iterations instead of 3 (ablation) | APVR · LongVideoBench | Avg accuracy: 64.9 -> 64.5 | No clear effect |
| add Pivot Frame Retrieval and Pivot Token Retrieval (ablation) | Qwen2.5-VL-7B · LongVideoBench | accuracy: 59.5 -> 64.9 | Improved |
| add APVR to Qwen2-VL-7B | Qwen2-VL-7B · MLVU dev | accuracy: 66.9 -> 73.4 | Improved |
| add APVR to VideoLLaMA3-7B | VideoLLaMA3-7B · VideoMME overall | accuracy: 66.2 -> 68.1 | Improved |
Closest papers by meaning
- Video-RAG: Visually-aligned Retrieval-Augmented Long Video Comprehension (NeurIPS 2025 · similarity 0.90)
- AdaVideoRAG: Omni-Contextual Adaptive Retrieval-Augmented Efficient Long Video Understanding (NeurIPS 2025 · similarity 0.90)
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.89)
- MUVR: A Multi-Modal Untrimmed Video Retrieval Benchmark with Multi-Level Visual Correspondence (NeurIPS 2025 · similarity 0.89)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.89)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 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.
