Video Motion Perception for Self-supervised Representation Learning

The motion of a video contains two factors: magnitude and direction, but most of the existing video self-supervised methods ignored the motion direction information. In this paper, we propose a Video Motion Perception (VMP) self-supervised framework, simultaneously taking account of the above two key factors. Specifically, a Motion Direction Perception Module (MDPM) is applied to asking the network to predict the moving direction of the video objects by using two well-designed handcraft strategies. Additionally, we analyze the characteristic of video motion in natural scenes and propose the Motion Change Perception Module (MCPM) accordingly for motion magnitude learning. Experimental results show that VMP achieves competitive performance on different benchmarks, including action recognition, video retrieval, and action similarity labeling.