Title: GenSpan: Leveraging Generation-Calibrated Motion Span Priors for Retrieving Moments in Multi-Verb Video Corpora

Abstract:

Video Corpus Moment Retrieval (VCMR) involves identifying both the correct video clip and its specific temporal segment that matches a natural-language query. This task is particularly difficult when dealing with multi-verb queries, as the precise ordering of temporal actions is crucial. Current methods frequently depend exclusively on textual data or static imagery, which limits their ability to grasp implicit motion dynamics. Consequently, these approaches often suffer from retrieval inaccuracies and temporal misalignments.

To address these limitations, we introduce GenSpan, a novel VCMR framework that utilizes generation-calibrated priors. GenSpan creates short auxiliary videos derived from LLM-selected subtitle cues and decomposed sub-events. Rather than treating these generated videos as direct retrieval targets, the model uses them as temporal priors. The framework employs a token selector to filter candidate-video features that align with the generated motion, followed by a bidirectional state-space model that efficiently predicts video-moment tuples.

Evaluations conducted on the TVR and ActivityNet-Captions datasets reveal that GenSpan enhances both corpus-level retrieval and moment localization. These improvements are especially notable for complex, multi-action queries. Furthermore, the proposed method achieves these gains while lowering computational costs in comparison to leading state-of-the-art multimodal baselines.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC