ACSNet: Action-Context Separation Network for Weakly Supervised Temporal Action Localization

The object of Weakly-supervised Temporal Action Localization (WS-TAL) is to localize all action instances in an untrimmed video with only video-level supervision. Due the lack frame-level annotations during training, current WS-TAL methods rely on attention mechanisms foreground snippets or frames that contribute classification task. This strategy frequently confuse context actual action, localization result. Separating and a core problem for precise WS-TAL, but it very challenging has been largely ignored literature. In this paper, we introduce Action-Context Separation Network (ACSNet) explicitly takes into account accurate localization. It consists two branches (i.e., Foreground-Background branch branch). first distinguishes from background within entire while further separates as context. We associate latent components positive component negative component), their different combinations can effectively characterize foreground, Furthermore, extended labels auxiliary categories facilitate learning action-context separation. Experiments THUMOS14 ActivityNet v1.2/v1.3 datasets demonstrate ACSNet outperforms existing state-of-the-art by large margin.