Active Video Perception: Iterative Evidence Seeking for Agentic Long Video Understanding

TL;DR

This paper introduces Active Video Perception (AVP), an interactive framework for long video understanding that actively seeks relevant evidence, significantly improving accuracy and efficiency over existing methods.

Contribution

AVP is a novel evidence-seeking framework that enables LVU agents to actively decide observations, improving relevance and efficiency in long video comprehension.

Findings

AVP achieves state-of-the-art accuracy on five LVU benchmarks.

AVP reduces inference time by 81.6% compared to baseline methods.

AVP outperforms previous agentic approaches by 5.7% in average accuracy.

Abstract

Long video understanding (LVU) is challenging because answering real-world queries often depends on sparse, temporally dispersed cues buried in hours of mostly redundant and irrelevant content. While agentic pipelines improve video reasoning capabilities, prevailing frameworks rely on a query-agnostic captioner to perceive video information, which wastes computation on irrelevant content and blurs fine-grained temporal and spatial information. Motivated by active perception theory, we argue that LVU agents should actively decide what, when, and where to observe, and continuously assess whether the current observation is sufficient to answer the query. We present Active Video Perception (AVP), an evidence-seeking framework that treats the video as an interactive environment and acquires compact, queryrelevant evidence directly from pixels. Concretely, AVP runs an iterative plan-observe-reflect process with MLLM agents. In each round, a planner proposes targeted video interactions, an observer executes them to extract time-stamped evidence, and a reflector evaluates the sufficiency of the evidence for the query, either halting with an answer or triggering further observation. Across five LVU benchmarks, AVP achieves highest performance with significant improvements. Notably, AVP outperforms the best agentic method by 5.7% in average accuracy while only requires 18.4% inference time and 12.4% input tokens.