Improving pseudo-labeling with reliable inter-camera distance encouragement for unsupervised person re-identification

Yiyu Chen, Zheyi Fan, Shuni Chen, Yixuan Zhu

Abstract: Unsupervised person re-identification (re-ID) aims to train a discriminative model without identity annotations. State-of-the-art methods usually follow a clustering-based pipeline, which utilizes a clustering algorithm to generate pseudo-labels and performs model training iteratively. Despite the success of these methods, camera variation has not been well considered in the distance calculation step to reduce noise labels, which is the main reason for noise label generation. To address this problem, we design a plug-and-play method with low computational complexity to improve the distance calculation step in the pseudo-label generation stage. Specifically, (1) to alleviate the influence of intra-camera samples and select reliable inter-camera pairs, we propose inter-camera k-reciprocal nearest neighbor (IKNN) to mine reliable inter-camera positive pairs; and (2) to merge the relationship of reliable inter-camera pairs into the distance for better clustering, we design adaptive inter-camera encouragement (AIE) to encourage the distance of inter-camera positive pairs. Extensive experiments show consistent improvements in comparison with various state-of-the-art unsupervised re-ID methods, demonstrating the wide applicability and effectiveness of our method.