Attention-guided Boundary Refinement on Anchor-free Temporal Action Detection

Henglin Shi, Haoyu Chen, Guoying Zhao

Abstract: Modelling temporal dependencies is important for accurate action detection. In this work, we develop a temporal attention unit to mine the global dependencies among features from different temporal locations. Additionally, based on the developed temporal attention unit, we propose an attention-guided boundary refinement module for revising action prediction results. Besides, we integrate the proposed module into a contemporary anchor-free detector for performing temporal action detection. To evaluate the proposed method, experiments are carried out on two large-scale temporal action detection datasets, namely THUMOS14 and ActivityNet1.3 datasets. Experimental results show that the action detection performance is significantly boosted by the proposed temporal attention module which outperforms several state-of-the-art methods.

External IDs:dblp:conf/scia/ShiCZ23