Research map / Video / Video-language understanding
QSVideo: Query-Conditioned Semantic Temporal Retrieval for Video Understanding
Wei Ao, Lan Wang, Vishnu Boddeti
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 QSVideo retrieval to MiniCPM-o 2.6 (16 frames) | MiniCPM-o 2.6 · LVBench | overall accuracy (%): 38.9 -> 45.8 | Improved |
| add QSVideo retrieval to MiniCPM-o 2.6 (32 frames) | MiniCPM-o 2.6 · LVBench | overall accuracy (%): 42.0 -> 49.1 | Improved |
| add QSVideo retrieval to MiniCPM-o 2.6 (8 frames) | MiniCPM-o 2.6 · StreamingBench real-time | overall accuracy (%): 64.44 -> 77.68 | Improved |
| add QSVideo retrieval to MiniCPM-o 2.6 (16 frames) | MiniCPM-o 2.6 · StreamingBench real-time | overall accuracy (%): 65.12 -> 79.36 | Improved |
| plug QSVideo into MiniCPM-V 2.6 at 8 frames | MiniCPM-V 2.6 · StreamingBench real-time | overall accuracy (%): 66.36 -> 70.96 | Improved |
| plug QSVideo into LLaVA-OV, gain shrinks with more frames | LLaVA-OV · StreamingBench real-time | overall accuracy (%): 72.12 -> 74.48 (8 frames); 73.56 -> 74.40 (32 frames) | Improved |
| replace holistic single-score ranker with structured object/action/location QSRanker (ablation) | QSRanker · LVBench ranking Recall@100 | Recall@100 (%): 52.78 -> 58.20 | Improved |
| replace holistic ranker with QSRanker at top-1 retrieval (ablation) | QSRanker · LVBench ranking Recall@1 | Recall@1 (%): 17.38 -> 17.51 | No clear effect |
| increase lambda (relevance weight) from 0.1 to 0.9 at 16 frames (ablation) | QSRetrieval · LVBench (16 frames) | overall accuracy (%): 46.1 (0.1) vs 45.8 (0.9) | No clear effect |
| increase lambda from 0.1 to 0.9 at 32 frames (ablation) | QSRetrieval · LVBench (32 frames) | overall accuracy (%): 48.2 -> 49.1 | Improved |
| use batch inference and data parallelism for semantic ranker (ablation) | QSRanker · latency breakdown | latency (s): 608.40s -> 61.32s | Improved |
| use QSVideo with 8B VLM instead of SEAL 34B | QSVideo · LVBench end-to-end latency | latency (s): 1702.24 -> 88.73 | Improved |
Closest papers by meaning
- Video-RAG: Visually-aligned Retrieval-Augmented Long Video Comprehension (NeurIPS 2025 · similarity 0.90)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.90)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.90)
- AdaVideoRAG: Omni-Contextual Adaptive Retrieval-Augmented Efficient Long Video Understanding (NeurIPS 2025 · similarity 0.90)
- Vgent: Graph-based Retrieval-Reasoning-Augmented Generation For Long Video Understanding (NeurIPS 2025 · similarity 0.90)
- MUVR: A Multi-Modal Untrimmed Video Retrieval Benchmark with Multi-Level Visual Correspondence (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.
