Video Action Detection Based On Deep Learning

Video action detection aims to find temporal action regions in videos.Generally,according to the information used,existing works can be categorized into two groups:local information based and global information based.However,both of them did not consider the global and local information jointly,which can be promising to generated proposals with high overlap rate and recall rate.To retrieve high quality proposals,we propose the Proposal Correction Network(PCN)to correct the candidate proposals generated by local information,which contains Candi-date Proposal Generation and Proposal Correction.When correcting the candidate pro-posals,2D temporal feature map is introduced in PCN to train the candidate proposals generation and proposals correction jointly and generate proposal boundary correction map.Moreover,a new Temporal Distance-based Generalized IoU(TDIoU)is introduce to address the problem of unbalanced number of positive or negative samples.We conduct experiments on two challenging datasets:Activity Net-1.3 and THUMOS14,where PCN significantly outperforms state-of-the-art temporal action proposal generation methods.