Video-based vehicle re-identification via channel decomposition saliency region network

摘要

Vehicle re-identification has become an important research topic for its application prospect in real-world, such as intelligent security system and intelligent traffic management. The current vehicle re-identification algorithms mainly run on image-based datasets, while video-based datasets are very rare in the community. Therefore we collect a dataset named as Veri-Video-763, including 763 vehicle IDs and 5828 tracks. In addition, we propose a channel decomposition saliency region network, including three modules to improve the video-based vehicle re-identification. The channel decomposition saliency region extraction (CDSRE) module generate significant masks to detect multiple significant local regions by channel decomposition. The global-local stacking module encode the convolutional features of the salient regions and the global pooling feature together into re-identification feature vectors. The distributed symmetric sampling (DSS) module propose a novel video clip sampling algorithm to improve the unity and difference of the video clips. Extensive experiments demonstrate the effectiveness of our proposed methods, and thus can be considered as one strong baseline. Dataset and code are available on https://github.com/wyf27/Veri-Video-763.