Camera-irrelevant Feature Invariance Learning for Unsupervised Person Re-identification

Abstract Unsupervised person re-identification (re-ID) aims to match pictures of target pedestrians from some other cameras without any labels. The existing methods usually adopt a cluster-based pipeline generate false labels for training. However, because the differences in camera configuration, taking Angle, illumination intensity and pedestrian background, unlabeled training samples captured by different are hard classify into cluster. To solve this problem, paper proposes camera-irrelevant feature learning framework, which learns features distances through network clustering. Specifically, during training, images styles generated, enriches cross-camera learning. During clustering, we propose distance matrix, reduces inter-class variance caused variance.