Graph-Induced Contrastive Learning for Intra-Camera Supervised Person Re-Identification

Abstract

Intra-camera supervision (ICS) for person re-identification (Re-ID) assumes that identity labels are independently annotated within each camera view and no inter-camera identity association is labeled. It is a new setting proposed recently to reduce the burden of annotation while expect to maintain desirable Re-ID performance. However, the lack of inter-camera labels makes the ICS Re-ID problem much more challenging than the fully supervised counterpart. By investigating the characteristics of ICS, this paper proposes a graph-induced contrastive learning (GCL) approach to address this issue. More specifically, we first formulate the cross-camera ID association task as a graph partitioning problem subjected to ICS-specific constraints and design a greedy agglomeration algorithm to solve it. Then, we propose a graph-induced contrastive loss that unifies both intra-and inter-camera learning into a contrastive learning framework to learn a Re-ID model. The cross-camera ID association step and the Re-ID model contrastive learning step are alternatively iterated, by which we progressively obtain a highly discriminative Re-ID model. Extensive experiments on three large-scale datasets show that our approach outperforms all previous ICS works. Especially, it gains 15.7% Rank-1 and 14.3% mAP improvements on the challenging MSMT17 dataset. Moreover, our approach performs even comparable to state-of-the-art fully supervised methods on all of the three datasets.

Anti‐occlusion person re‐identification via body topology information restoration and similarity evaluation

“…Existing research in this field has predominantly focused on three key aspects: learning robust representations of pedestrian images [21,22], measuring similarity between learnt features [1,[23][24][25], and re-ranking optimisation [26][27][28], all aimed at predicting whether two images correspond to the same person. Most ReID methods currently address the challenge of holistic pedestrian image matching [29][30][31][32][33]. Deng et al [34] implemented ReID in a similarity preserving generative adversarial network, which consists of Siamese network and CycleGAN, through two constraints: self-similarity of the images before and after translation, and domain dissimilarity of the translated source and a target image.…”

“…Existing research in this field has predominantly focused on three key aspects: learning robust representations of pedestrian images [21, 22], measuring similarity between learnt features [1, 23–25], and re‐ranking optimisation [26–28], all aimed at predicting whether two images correspond to the same person. Most ReID methods currently address the challenge of holistic pedestrian image matching [29–33]. Deng et al.…”

Offline-Online Associated Camera-Aware Proxies for Unsupervised Person Re-identification

“…In this subsection, we analyze the sensitivity of the hyperparameters involved in O2CAP. Among all hyper-parameters, the memory updating rate µ and the temperature factor τ have been investigated in many other works so that we simply follow [4], [11] to set them. Here, we conduct experiments on DukeMTMC-reID and MSMT17 to investigate the sensitivity of the remaining hyper-parameters, which include the number of hard negative proxies (K 1 ), the number of positive proxies associated online (K 2 ), and the weight w in the instance-proxy balanced similarity.…”

“…State-of-the-art performance is achieved mostly by supervised methods [1], [2], requiring full labels that are expensive and time-consuming to annotate. Recently, semi-supervised [3], [4] and unsupervised [5], [6] Re-ID have been attracting more and more research interest, in a hope to reduce annotation cost and make the techniques more practical to real-world deployments. Although considerable progress has been achieved in these tasks, there is still a big gap in performance compared to the supervised counterparts.…”

A Generative Approach to Person Reidentification

“…Metric Learning is typically based on some form of contrastive learning [ 24 ], either using contrastive loss [ 25 , 26 , 27 , 28 ] or, more frequently, triplet loss [ 29 , 30 , 31 , 32 ]. Both approaches aim to encourage similar examples to be close in the feature space and dissimilar examples to be separated.…”