p-TSM: Learning to shift temporal features point-wise for action recognitionDownload PDF

09 Nov 2022 (modified: 09 Nov 2022)OpenReview Archive Direct UploadReaders: Everyone
Abstract: We propose a method of temporal shifting of features for video image recognition. Compared to the TSM module whose shifting channels are fixed and spatially uniform, the amount of temporal shift is learnt and spatially variable. The effectiveness of the proposed method is demonstrated by comparing the network model with the proposed shift module with the conventional method.
0 Replies

Loading