基于时序一致性调整网络的动作分割方法OA

Objectives Aiming at the problems of insufficient recognition of tail actions caused by long-tail distribution in temporal action segmentation,as well as mis-segmentation and over-segmentation resulting from existing methods that tend to ignore temporal dependencies between actions.Methods A temporal action segmentation method based on a temporal consistency adjustment network is proposed.Built upon an action segmentation backbone,the method constructs a temporal exchange self-supervised module to enhance the recognition ability of tail actions,designs a dynamic temporal adjustment module to learn the dependencies between actions,and introduces a segmental nearest class mean module to alleviate over-segmentation.Results Experimental results on the 50Salads and Breakfast datasets show that the proposed method achieves certain improvements on most metrics.Conclusions The method can effectively address the issues of long-tail distribution and temporal consistency.