Guohui Ding, Jing Li, Yimin Xu, Rui Hong Zhou · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3836226
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).
Cross-modal retrieval is a fundamental task in multimedia understanding, which aims to retrieve samples in one modality that share the same or similar semantics by querying with another modality. Existing methods typically treat labels merely as supervisory signals or use label information to enforce semantic consistency between cross-modal features, while overlooking the high-order semantic information implicit in the relationships among multiple labels and their combinations. To explore accordingly, this paper proposes a novel scene-aware cross-modal retrieval method. This method combines multiple single tags into a tag group as a scene and explicitly models the higher-order semantic relationships guided by the scene. Furthermore, to mitigate the interference of background noise, a self-attention mechanism is introduced within each scene to dynamically learn the semantic contribution weights of different cross-modal samples under the same scene. Meanwhile, considering that different scenes have varying importance across the entire dataset, we further leverage global label co-occurrence statistics to assign overall weights to scenes, so as to measure the relative importance of different scenes at the global level. Extensive experimental results demonstrate that the proposed method achieves significantly superior performance compared with existing approaches on three widely used cross-modal retrieval benchmark datasets.
No comments yet — start the discussion below.