Meta Distribution Alignment for Generalizable Person Re-Identification

Domain Generalizable (DG) person ReID is a challenging task which trains a model on source domains yet generalizes well on target domains. Existing methods use source domains to learn domain-invariant features, and assume those features are also irrelevant with target domains. However, they do not consider the target domain information which is unavailable in the training phrase of DG. To address this issue, we propose a novel Meta Distribution Alignment (MDA) method to enable them to share similar distribution in a test-time-training fashion. Specifically, since high-dimensional features are difficult to constrain with a known simple distribution, we first introduce an intermediate latent space constrained to a known prior distribution. The source domain data is mapped to this latent space and then reconstructed back. A meta-learning strategy is introduced to facilitate generalization and support fast adaption. To reduce their discrepancy, we further propose a test-time adaptive updating strategy based on the latent space which efficiently adapts model to unseen domains with a few samples. Extensive experimental results show that our model outperforms the state-of-the-art methods by up to 5.2% R-1 on average on the large-scale and 4.7% R-1 on the single-source domain generalization ReID benchmark. Source code is publicly available at https://github.com/haoni0812/MDA.git.

1. Introduction

Person Reidentification (ReID) aims to match persons with the same ID across different camera views. Thanks to the development of deep convolutional neural networks (CNNs) [14], supervised ReID and unsupervised domain adaptation (UDA) [42] have achieved remarkable performance. However, they both need data of target domain for training. In real-world applications, the ReID system will inevitably search persons in unseen domains. Therefore, domain generalization (DG) ReID has attracted extensive research attention in a practical setting. Figure 1.

Illustration of our idea. Since target domain data (green) is unavailable during training, we cannot directly align source and target distributions. To address this issue, we align them to a prior distribution (purple) during training (source domains) and testing (target domains). Considering that high-dimensional ID features are difficult to constrain to a prior distribution, ID features are encoded into latent embedding space. The same prior distribution and decoder guarantee the same generated feature distribution.