LSTIF:Long-short Temporal Information Fusion Architecture for Video-based Person Re-identification

Person re-identification is a major application of computer vision in reality. Since the data obtained by monitoring in real life is often in video format, and the walking poses of pedestrians are different, in addition to the appearance of pedestrians, how to obtain the motion features of pedestrians, is extremely important for video-based person re-identification. Therefore, for the temporal information of the video, we propose a Long-short Temporal Information Fusion (LSTIF) network. We aggregate temporal information from two perspectives, short-term features containing detailed information and long-term features containing global information. Simultaneously, in order to reduce the amount of calculation, this network also uses non-local blocks, and extend the outpu feature map to the same size as the input, which is convenient for calculation. This paper verifies the effectiveness of our method on two commonly used datasets iLIDS-VID and DukeMTMC-VideoReID.