Unsupervised Video-based Person Re-identification Based on The Joint Global-local Metrics

Abstract

At present, supervised video-based person re-identification has achieved excellent performance. However, the initial video data obtained from real scenes are often unlabeled. Labelling such data is very time-consuming. If unsupervised learning can be effectively applied to these data, so much cost will be saved. In this paper, based on the joint global and local metric, an unsupervised video-based person re-identification method is proposed, which takes both the global information of a video sequence and the local information between the video frames into account to better distinguish different appearances of the same pedestrian. The global similarity and local similarity are calculated using global and local features, respectively. Meanwhile, a diversity constraint is used as an aid for cluster merging and evaluation. In the training process, the network is optimized by combining cluster mutual exclusion loss and center loss, which reduces the within-class differences and enlarges the between-class differences. Experiments on two benchmark datasets, MARS and DukMTMC-VideoReID, the results show that this method has higher accuracy and stabilityshow that the proposed method can achieve higher accuracy and is more stable than most state-of-the-art unsupervised methods.