Global context-aware attention model for weakly-supervised temporal action localization

Temporal action localization (TAL) is a significant and challenging task in the field of video understanding. It aims to locate the start and end timestamps of the actions in a video and recognize their categories. However, efficient action localization often requires extensive precise annotations. Therefore, the researchers propose weakly-supervised temporal action localization (WTAL), which aims to locate action instances in a video using only video level annotations. The existing WTAL methods lack the ability to distinguish the action context information effectively, including the pre-action and post-action scenes, which blur the action boundary and lead to the inaccurate action location. To solve the above problems, this paper proposes a global context-aware attention model (GCAM). Firstly, GCAM designs the mask attention module (MAM) to restrict the model's receptive field and make the model focus on localized features related to the action context. It enhances the ability to distinguish the action context information and clearly locate the start and end timestamps of the actions. Secondly, GCAM introduces the context broadcasting module (CBM), which supplements the global context information to keep the features intact in temporal dimension. This module solves the issue that the model overemphasizes the localized features due to the addition of the MAM. Extensive experiments on the THUMOS14 and ActivityNet1.2 datasets demonstrate the effectiveness of GCAM. On the THUMOS14 dataset, GCAM achieves an average mean average precision (mAP) of 49.5 %, representing a 2.2 % improvement over existing WTAL methods. On the ActivityNet1.2 dataset, GCAM achieves an average mAP of 27.2 %, representing a 0.3 % improvement over existing WTAL methods. These results highlight the superior performance of GCAM in accurately localizing actions in videos. • A global context-aware attention with mask for action boundary clearing by limiting receptive field and localized features. • A context broadcasting solves segment loss from overemphasis on localized features by using cross-frame associations. • Achieves improvements on two datasets compared with the state-of-the-art methods.