Disentanglement-based Cross-Domain Feature Augmentation for Effective Unsupervised Domain Adaptive Person Re-identification

Abstract

Unsupervised domain adaptive (UDA) person reidentification (ReID) aims to transfer the knowledge from the labeled source domain to the unlabeled target domain for person matching. One challenge is how to generate target domain samples with reliable labels for training. To address this problem, we propose a Disentanglement-based Cross-Domain Feature Augmentation (DCDFA) strategy, where the augmented features characterize well the target and source domain data distributions while inheriting reliable identity labels. Particularly, we disentangle each sample feature into a robust domain-invariant/shared feature and a domain-specific feature, and perform cross-domain feature recomposition to enhance the diversity of samples used in the training, with the constraints of cross-domain ReID loss and domain classification loss. Each recomposed feature, obtained based on the domain-invariant feature (which enables a reliable inheritance of identity) and an enhancement from a domain specific feature (which enables the approximation of real distributions), is thus an "ideal" augmentation. Extensive experimental results demonstrate the effectiveness of our method, which achieves the state-ofthe-art performance.

Calibrated Feature Decomposition for Generalizable Person Re-Identification

“…To address this problem, several tailored DG-ReID methods [6,7,25,28,31,32,43,53] have been proposed, which can be mainly divided into three categories: metalearning based model, ensemble learning based model and disentanglement based model. Due to the success of disentangled learning, the DG-ReID methods based on disentangled learning [10,20,41,54] improve the model generalization ability by disentangling person representations into identity-irrelevant interference and id-invariant feature. Specifically, Eom et al [10] propose to disentangle identityrelated and -unrelated features from person images and adopt identity shuffle GAN to enhance the person representation.…”

“…Specifically, Eom et al [10] propose to disentangle identityrelated and -unrelated features from person images and adopt identity shuffle GAN to enhance the person representation. Zhang et al [41] propose a Disentanglement-based Cross-Domain Feature Augmentation (DCDFA) strategy to generate virtual samples in the feature space by adding disentangled domain-specific enhancements upon disentangled domain-shared identity bases. Zhang et al [40] construct a structural causal model (SCM) to disentangle the person representation into identity-specific factors and domain-specific factors.…”

“…Then, this calibrated person representation F is decomposed into two parts: soft entangled feature R ∈ R h×w×c and pure domain-relevant feature R − ∈ R h×w×c (i.e., this feature contains the domain information that is irrelevant to id information). To adequately filter out this pure domain-relevant feature R − , following the previous disentanglement-based ReID methods [20,41], we utilize learnable channel attention vector a…”

“…Compared to the two types of aforementioned methods, as illustrated in Fig. 1 (a), the disentanglement based model is a more straightforward and efficient approach for DG-ReID [10,20,40,41,54]. It aims to directly disentangle identity-irrelevant interference and domaininvariant feature from the learned representations, and employ the domain-invariant feature to improve the generalization.…”

“…and pure id-relevant feature R + from the residual feature F. Here, we study the influence of different disentanglement designs. Following conventional disentanglement methods[10,20,40,41], previous feature decomposition (PFD) methods directly disentangle the input feature F into id-relevant feature R − and id-irrelevant/domainrelated feature R + . Tab.…”