Progressive Cross-Camera Soft-Label Learning for Semi-Supervised Person Re-Identification

Abstract

In this paper, we focus on the semi-supervised person re-identification (Re-ID) case, which only has the intracamera (within-camera) labels but not inter-camera (crosscamera) labels. In real-world applications, these intra-camera labels can be readily captured by tracking algorithms or few manual annotations, when compared with cross-camera labels. In this case, it is very difficult to explore the relationships between cross-camera persons in the training stage due to the lack of cross-camera label information. To deal with this issue, we propose a novel Progressive Cross-camera Soft-label Learning (PCSL) framework for the semi-supervised person Re-ID task, which can generate cross-camera soft-labels and utilize them to optimize the network. Concretely, we calculate an affinity matrix based on person-level features and adapt them to produce the similarities between cross-camera persons (i.e., cross-camera soft-labels). To exploit these soft-labels to train the network, we investigate the weighted cross-entropy loss and the weighted triplet loss from the classification and discrimination perspectives, respectively. Particularly, the proposed framework alternately generates progressive cross-camera soft-labels and gradually improves feature representations in the whole learning course. Extensive experiments on five large-scale benchmark datasets show that PCSL significantly outperforms the state-of-the-art unsupervised methods that employ labeled source domains or the images generated by the GAN-based models. Furthermore, the proposed method even has a competitive performance with respect to deep supervised Re-ID methods.

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

“…Finally, we compare the proposed method with existing intra-camera supervised methods, including MTML [16], PCSL [17], ACAN [23], and MATE [22]. Table 3 shows that our graph-induced contrastive learning (GCL) approach consistently outperforms these methods on all datasets.…”

“…This work aims to learn a person Re-ID model under intracamera supervision (ICS), which is a new semi-supervised setting proposed very recently [16], [17]. It assumes that identity labels are independently annotated within each cam-era view and no inter-camera identity association is labeled.…”

“…• We propose a graph partitioning based strategy for ID association in ICS. In contrast to previous ICS works that use soft-labeling [17] or multi-label learning [16], [22] strategies, our association strategy can ensure the ICS-specific constraints to be satisfied so that more reliable association results are obtained. • We propose a graph-induced contrastive learning method to learn the Re-ID model.…”

“…• We propose a graph-induced contrastive learning method to learn the Re-ID model. In contrast to previous ICS works [16], [17], [22], [23] that treat intraand inter-camera learning differently and separately, our approach unifies both learning parts into the same framework and performs them end-to-end, enabling the learning more effective.…”

“…Existing ICS methods address the Re-ID problem by considering both intra-and inter-camera learning perspectives. In supervised intra-camera learning, PCSL [17] and ACAN [23] adopt the extensively used triplet loss [2], MTML [16] and MATE [22] construct a multi-branch structure to learn parametric classifiers, and Precise-ICS [24] learns cameraspecific non-parametric classifiers. In unsupervised intercamera learning, ACAN [23] develops a multi-camera adversarial learning approach to reduce the cross-camera data distribution discrepancy, PCSL [17] utilizes a soft-labeling scheme, MTML [16] and MATE [22] adopt multi-label learning strategies, and Precise-ICS [24] trains a parametric Re-ID model with the supervision of pseudo labels generated by graph association.…”