Unsupervised dual-branch cross-domain person re-identification based on domain-invariant features

Abstract

To address the distribution discrepancy between the source and target domains caused by camera-specific style variations, we propose a unsupervised dual-branch cross-domain person re-identification framework based on domain-invariant feature learning. Specifically, during the source domain pre-training phase, considering the distribution shift induced by inter-camera style differences, we treat each camera as an independent style domain. CycleGAN is employed to perform camera-style transfer, which significantly enhances the diversity of training samples and alleviates inter-domain distribution bias. To simultaneously capture fine-grained local details and high-level semantic context, we place the IBN and Non-local modules in Layer2 and Layer3 of the network. Additionally, a fixed exponent GeM pooling strategy is adopted to improve both the discriminability and generalizability of the learned features. During the target domain adaptation stage, in order to suppress the noise introduced by clustering-generated pseudo labels, a dual-branch symmetric architecture is constructed. An Exponential Moving Average model is maintained to generate soft pseudo labels. Using complementary supervision between hard and soft labels, our method effectively mitigates label noise and enhances robustness. Extensive experiments conducted on three widely used datasets (Market1501, DukeMTMC-reID, and MSMT17) demonstrate the effectiveness of the proposed method in both unsupervised domain adaptation and purely unsupervised person re-identification tasks.