EventSTU: Event-Guided Efficient Spatio-Temporal Understanding for Video Large Language Models

TL;DR

EventSTU introduces an event-guided, training-free framework for efficient video understanding that reduces computational costs by exploiting event-based cues for spatio-temporal sampling and pruning, supported by a new multimodal benchmark.

Contribution

It proposes a novel event-guided, training-free approach for efficient spatio-temporal video understanding, including algorithms for keyframe sampling and token pruning, and introduces the EventBench benchmark.

Findings

Achieves 3.01x FLOPs reduction

Attains 3.10x speedup in prefilling

Improves performance over baseline

Abstract

Video large language models have demonstrated strong video understanding capabilities but suffer from high inference costs due to the massive number of tokens in long videos. Inspired by event-based vision, we propose an event-guided, training-free framework for efficient spatio-temporal understanding, named EventSTU. In the temporal domain, we design a coarse-to-fine keyframe sampling algorithm that exploits the change-triggered property of event cameras to eliminate redundant frames. In the spatial domain, we design an adaptive token pruning algorithm that leverages the visual saliency of events as a zero-cost prior to guide spatial reduction. From a holistic spatio-temporal perspective, we further integrate question relevance from keyframe sampling to adaptively allocate token pruning budgets. To facilitate evaluation, we construct EventBench, the first event-inclusive, human-annotated multimodal benchmark that covers diverse real-world scenarios. Beyond physical event cameras, EventSTU also supports general video understanding using simulated events. Comprehensive experiments show that EventSTU achieves 3.01x FLOPs reduction and 3.10x prefilling speedup over the strongest baseline while still improving performance.