Research map / Video / Video-language understanding
MDP3: A Training-free Approach for List-wise Frame Selection in Video-LLMs
Hui Sun, Shiyin Lu, Huanyu Wang, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Ming Li
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 |
|---|---|---|---|
| replace uniform sampling with mDP3 frame selection | VILA-V1.5-8B · Video-MME (w/o subs, 8 frames) | accuracy: 47.5 -> 53.3 | Improved |
| replace uniform sampling with mDP3 frame selection | MiniCPM-V2.6-7B · Video-MME (w/o subs, 8 frames) | accuracy: 52.6 -> 58.0 | Improved |
| replace uniform sampling with mDP3 frame selection | LLaVA-OneVision-7B · Video-MME (w/o subs, 8 frames) | accuracy: 53.6 -> 59.6 | Improved |
| replace uniform sampling with mDP3 frame selection | Ovis2-7B · Video-MME (w/o subs, 8 frames) | accuracy: 58.9 -> 63.9 | Improved |
| add mDP3 instead of Frame-Voyager | VILA-V1.5-8B · Video-MME (w/o subs, 8 frames) | accuracy: 50.5 -> 53.3 | Improved |
| add sequentiality (Markov segment-wise selection) to DPP (ablation) | mDP3 on VILA-V1.5-8B · Video-MME (w/o subs, 8 frames) | accuracy: 51.8 -> 53.3 | Improved |
| use conditional kernel in RKHS instead of cosine similarity (ablation) | mDP3 on VILA-V1.5-8B · Video-MME (w/o subs, 8 frames) | accuracy: 50.2 -> 53.3 | Improved |
| use conditional multi-gaussian kernel instead of unconditional MGK (ablation) | mDP3 on VILA-V1.5-8B · Video-MME (w/o subs, 8 frames) | accuracy: 48.9 -> 53.3 | Improved |
| add diversity and sequentiality on top of SigLIP query-relevance top-k (ablation) | mDP3 on VILA-V1.5-8B · Video-MME (w/o subs, 8 frames) | accuracy: 50.6 -> 53.3 | Improved |
| use rule-based shot boundary selection (RGB histogram) instead of uniform (ablation) | VILA-V1.5-8B · Video-MME (w/o subs, 8 frames) | accuracy: 47.5 -> 45.9 | Hurt |
| use Edges Change Ratio selection instead of uniform (ablation) | VILA-V1.5-8B · Video-MME (w/o subs, 8 frames) | accuracy: 47.5 -> 47.3 | No clear effect |
| mDP3 on MiniCPM-V2.6 for long videos | MiniCPM-V2.6-7B · Video-MME long (w/o subs) | accuracy: 46.4 -> 48.4 | Improved |
Later papers that compared with this method
1 later accepted paper reports MDP3 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Towards Sparse Video Understanding and Reasoning (CVPR 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Frame-Voyager — Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- M-LLM Based Video Frame Selection for Efficient Video Understanding (CVPR 2025 · similarity 0.92)
- 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)
- Recurrent Attention-based Token Selection for Efficient Streaming Video-LLMs (NeurIPS 2025 · similarity 0.88)
- Flexible Frame Selection for Efficient Video Reasoning (CVPR 2025 · similarity 0.88)
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.88)
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.
