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
TL;DR
This work tackles frame selection for Video-LLMs by enforcing three core principles: query relevance, list-wise diversity, and sequentiality. It introduces mDP3, a training-free, model-agnostic framework that combines a conditional Gaussian kernel in RKHS (CMGK) to measure query-conditioned frame similarity, determinantal point processes (DPP) to capture diversity and relevance, and a Markov decision process (MDP) to allocate a fixed budget of frames across video segments for sequentiality. The method achieves a (1 - 1/e) approximation for the NP-hard list-wise selection problem with pseudo-polynomial complexity, and empirically demonstrates significant improvements when integrated with state-of-the-art Video-LLMs across three long-video benchmarks. The results support the practicality of training-free, plug-and-play frame selection to boost video understanding while reducing input frames and computational costs. The work also outlines future directions, including adaptive selection sizes and test-time computation strategies to scale further with minimal overhead.
Abstract
Video large language models (Video-LLMs) have made significant progress in understanding videos. However, processing multiple frames leads to lengthy visual token sequences, presenting challenges such as the limited context length cannot accommodate the entire video, and the inclusion of irrelevant frames hinders visual perception. Hence, effective frame selection is crucial. This paper emphasizes that frame selection should follow three key principles: query relevance, list-wise diversity, and sequentiality. Existing methods, such as uniform frame sampling and query-frame matching, do not capture all of these principles. Thus, we propose Markov decision determinantal point process with dynamic programming (MDP3) for frame selection, a training-free and model-agnostic method that can be seamlessly integrated into existing Video-LLMs. Our method first estimates frame similarities conditioned on the query using a conditional Gaussian kernel within the reproducing kernel Hilbert space~(RKHS). We then apply the determinantal point process~(DPP) to the similarity matrix to capture both query relevance and list-wise diversity. To incorporate sequentiality, we segment the video and apply DPP within each segment, conditioned on the preceding segment selection, modeled as a Markov decision process~(MDP) for allocating selection sizes across segments. Theoretically, MDP3 provides a \((1 - 1/e)\)-approximate solution to the NP-hard list-wise frame selection problem with pseudo-polynomial time complexity, demonstrating its efficiency. Empirically, MDP3 significantly outperforms existing methods, verifying its effectiveness and robustness.