Camera-Aware Proxies for Unsupervised Person Re-Identification

Abstract

This paper tackles the purely unsupervised person re-identification (Re-ID) problem that requires no annotations. Some previous methods adopt clustering techniques to generate pseudo labels and use the produced labels to train Re-ID models progressively. These methods are relatively simple but effective. However, most clustering-based methods take each cluster as a pseudo identity class, neglecting the large intra-ID variance caused mainly by the change of camera views. To address this issue, we propose to split each single cluster into multiple proxies and each proxy represents the instances coming from the same camera. These camera-aware proxies enable us to deal with large intra-ID variance and generate more reliable pseudo labels for learning. Based on the camera-aware proxies, we design both intra and inter-camera contrastive learning components for our Re-ID model to effectively learn the ID discrimination ability within and across cameras. Meanwhile, a proxy-balanced sampling strategy is also designed, which facilitates our learning further. Extensive experiments on three large-scale Re-ID datasets show that our proposed approach outperforms most unsupervised methods by a significant margin. Especially, on the challenging MSMT17 dataset, we gain 14.3 percent Rank-1 and 10.2 percent mAP improvements when compared to the second place. Code is available at: https://github.com/Terminator8758/CAP-master.

Discriminatively Unsupervised Learning Person Re-Identification via Considering Complicated Images

“…Purely unsupervised learning person re-ID. Purely unsupervised learning person re-ID methods completely train models on unlabeled data [ 5 , 6 , 7 , 9 , 10 , 24 , 25 , 26 , 27 , 28 ]. The training process usually consists of four main steps, including clustering to generate pseudo labels, constructing a memory dictionary, computing the contrastive loss, and updating the feature representations.…”

“…They treated each image as an individual cluster and gradually grouped similar images into one cluster to finally obtain a great balance between the diversity across clusters and similarity in clusters. Wang et al [ 6 ] exploited camera-aware proxies in each cluster to further distinguish the images from different cameras, and thus generate more reliable pseudo labels. Based on the camera-aware pseudo labels, they designed both intra-camera and inter-camera contrastive loss to enhance the model’s identity discrimination ability.…”

“…Purely unsupervised learning person re-identification (re-ID) refers to recognizing the same person appearing in images captured by different cameras without any groundtruth labels. With the rapid development of unsupervised representation learning [ 1 , 2 , 3 ], the performance of purely unsupervised learning person re-ID based on contrastive loss has been gradually improving [ 4 , 5 , 6 , 7 , 8 ]. On the other hand, purely unsupervised learning person re-ID does not require any labeled information or complex training processes, which is easy to implement and deploy in real-world application scenarios.…”

Unsupervised Gait Recognition with Selective Fusion

“…Short-term Unsupervised Re-ID: Most fully-unsupervised Re-ID methods estimate pseudo labels for sequences, which can be roughly categorized into the clustering-based and non-clusteringbased methods. Clustering-based methods [23], [24], [25], [26] first estimate a pseudo label for each sequence and train the network with sequence similarity. In contrast, non-clustering-based methods [7], [8] regard each image as a a class and use a nonparametric classifier to push each similar image closer and pull all other images further.…”

“…SpCL [6] introduces a self-paced learning strategy and memory bank, gradually making generated sample features closer to reliable cluster centroids. To alleviate the high intra-class variance inside a cluster caused by camera styles, CAP [24] proposes cross-camera proxy contrastive loss to pull instances near their own camera centroids in a cluster. ICE [25] further explores inter-instance relationships instead of using camera labels to compact the clusters with hard contrastive loss and soft instance consistency loss.…”