Hao Yang, Yibo Zhao, Luyang Dai, Chunjie Ma, Cong Liu, Zan Gao · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3835822
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Video Moment Retrieval (VMR) aims to locate specific temporal segments in untrimmed videos that semantically align with given language queries. Existing proposal-based methods suffer from cross-modal granularity mismatch and fail to fully exploit the distinct temporal cues embedded in the 2D proposal map. To address these limitations, we propose a novel framework named Learning with Temporal Coordinate Priors (LTCP) for video moment retrieval. LTCP integrates event-level semantics for high-quality proposal map initialization and explicitly models proposal relationships by leveraging the temporal cues along the coordinate axes. Specifically, we first design an event-level semantic guidance module that dynamically aggregates frame-level video features into event-level features and fuses them with multimodal features via bidirectional Mamba, yielding robust and context-rich initialization features for the subsequent proposal map. Then, we develop a temporal-coordinate axial modeling module that leverages decoupled 1D axial operations to isolate and model relationships between proposals with shared boundaries. Extensive evaluations on three benchmarks (ActivityNet-Captions, TACoS, and Ego4D-NLQ) demonstrate that LTCP achieves highly competitive and often state-of-the-art performance, validating the effectiveness of our method.
No comments yet — start the discussion below.