Association Loss and Self-Discovery Cross-Camera Anchors Detection for Unsupervised Video-Based Person Re-Identification

Abstract

With the continuous improvement and development of cameras network, surveillance video has become the data source of the column stream, which greatly promotes the development of cross-camera person re-identification (Re-ID). However, supervised learning requires a lot of effort to manually label cross-cameras pairwise training data, which is lack of scalability and practical in actual video surveillance because there is a lack of well-labeled pairs of positive and negative samples under each camera. For addressing these negative effects, we set judgment conditions by using the association ranking method to self-discover positive and negative track-lets pairs of anchors with none of the pairwise ID labels, thereby defining a triplet loss. In order to optimize association loss for learning effective discriminative feature, the triplet loss adds adaptive weights according to the degree of easy-hard samples to generate an Adaptive Weighted Conditional Triplet Loss. Besides, for increasing the accuracy of self-discovering cross-camera anchors independently, which means successfully mine mutually best-matched track-lets and merge them under cross-camera, we use the top-rank from the intra-camera ranking list as a self-matched query sample which can double verify the matched-degree between top-rank. And eventually, we establish a new Association Loss and Self-Discovery Learning (ALSL) model with a complete end-to-end manner. We use three standard datasets, PRID2011, iLIDS-VID and MARS, to train the model and the experimental results prove that ALSL rank-1 is better than some superior video-based unsupervised person Re-ID methods.