“In the realm of temporal action localization (TAL), a novel method has been proposed to address the limitations of existing weakly supervised approaches when faced with an increasing number of action categories per video. This method leverages single-frame supervision by employing the K-means clustering algorithm to group annotated frames that exhibit distinctive characteristics of the same action category. These frames, serving as exemplars, enhance action discrimination and provide supplementary guidance to snippet sequence modeling. The convex combination fusion mechanism further improves the consistency of action discrimination, leading to robust class activation sequences for precise classification and localization. Experiments on THUMOS14, GTEA, and BEOID datasets demonstrate that this method outperforms current single-frame annotation-based approaches, achieving superior localization performance.”