Camera-aware graph multi-domain adaptive learning for unsupervised person re-identification

Recently, unsupervised person re-identification (Re-ID) has gained much attention due to its important practical significance in real-world application scenarios without pairwise labeled data. A key challenge for unsupervised person Re-ID is learning discriminative and robust feature representations under cross-camera scene variation. Contrastive learning approaches treat unsupervised representation learning as a dictionary look-up task. However, existing methods ignore both intra- and inter-camera semantic associations during training. In this paper, we propose a novel unsupervised person Re-ID framework, Camera-Aware Graph Multi-Domain Adaptive Learning (CGMAL), which can conduct multi-domain feature transfer with semantic propagation for learning discriminative domain-invariant representations. Specifically, we treat each camera as a distinct domain and extract image samples from every camera domain to forma mini-batch. A heterogeneous graph is constructed for representing the relationships between all instances in a mini-batch. Then a Graph Convolutional Network (GCN) is employed to fuse the image samples into a unified space and implement promising semantic transfer for providing ideal feature representations. Subsequently, we construct the memory-based non-parametric contrastive loss to train the model. In particular, we design an adversarial training scheme for transferring the knowledge learned by GCN to the feature extractor. Experimental experiments on three benchmarks validate that our proposed approach is superior to the state-of-the-art unsupervised methods.