A dynamic predictive transformer with temporal relevance regression for action detection

Abstract

This paper introduces a novel transformer network tailored to skeleton-based action detection in untrimmed long video streams. Our approach centers around three innovative mechanisms that collectively enhance the network’s temporal analysis capabilities. First, a new predictive attention mechanism incorporates future frame data into the sequence analysis during the training phase. This mechanism addresses the essential issue of the current action detection models: incomplete temporal modeling in long action sequences, particularly for boundary frames that lie outside the network’s immediate temporal receptive field, while maintaining computational efficiency. Second, we integrate a new adaptive weighted temporal attention system that dynamically evaluates the importance of each frame within an action sequence. In contrast to the existing approaches, the proposed weighting strategy is both adaptive and interpretable, making it highly effective in handling long sequences with numerous non-informative frames. Third, the network incorporates an advanced regression technique. This approach independently identifies the start and end frames based on their relevance to different frames. Unlike existing homogeneous regression methods, the proposed regression method is heterogeneous and based on various temporal relationships, including those in future frames in actions, making it more effective for action detection. Extensive experiments on prominent untrimmed skeleton-based action datasets, PKU-MMD, OAD, and the Charade dataset demonstrate the effectiveness of this network.