GENERALIZABLE PERSON RE-IDENTIFICATION WITH-OUT DEMOGRAPHICS

Abstract

Domain generalizable person re-identification (DG-ReID) aims to learn a readyto-use domain-agnostic model directly for cross-dataset/domain evaluation, while current methods mainly explore the demographic information such as domain and/or camera labels for domain-invariant representation learning. However, the above-mentioned demographic information is not always accessible in practice due to privacy and security issues. In this paper, we consider the problem of person re-identification in a more general setting, i.e., domain generalizable person re-identification without demographics (DGWD-ReID). To address the underlying uncertainty of domain distribution, we introduce distributionally robust optimization (DRO) to learn robust person re-identification models that perform well on all possible data distributions within the uncertainty set without demographics. However, directly applying the popular Kullback-Leibler divergence constrained DRO (or KL-DRO) fails to generalize well under the distribution shifts in real-world scenarios, since the convex condition may not hold for overparameterized neural networks. Inspired by this, we analyze and reformulate the popular KL-DRO by applying the change-of-measure technique, and then propose a simple yet efficient approach, Unit-DRO, which minimizes the loss over a new dataset with hard samples up-weighted and other samples down-weighted. We perform extensive experiments on both domain generalizable and cross-domain person ReID tasks, and the empirical results show that Unit-DRO achieves superior performance compared to all baselines without using demographics.

1. INTRODUCTION

Person re-identification (ReID) aims to find the correspondences between person images from the same identity across multiple camera views. As illustrated in Figure 1 , previous studies mainly follow three different settings: 1) supervised person ReID Zhang et al. (2020) , where training and test data are independently and identically (i.i.d) drawn from the same distribution. Though recent supervised methods have achieved remarkable performance, they are usually non-robust in out-of-distribution (OOD) settings; 2) unsupervised domain adaptative person ReID (UDA-ReID) and cross-domain person ReID (CD-ReID) Luo et al. (2020) , where UDA-ReID relies on large amounts of unlabeled data for retraining and CD-ReID cannot exploit the benefits brought by multisource domains; 3) domain generalizable person ReID (DG-ReID) Dai et al. (2021a) , where the model is trained on multiple large-scale datasets and tested on unseen domains directly without extra data collection/annotation and model updating on new domains. Therefore, DG-ReID is receiving increasing attention due to its great value in real-world person retrieval applications. In this paper, we evaluate the proposed Unit-DRO for person ReID by comparing it with existing DG-ReID and CD-ReID methods. Unit-DRO outperforms a variety of recent methods with a large margin on both DG-ReID and CD-ReID benchmarks, even including those methods using demographics. To better understand the proposed Unit-DRO, we perform comprehensive ablation studies on several important components, such as the multi-step τ * solution and the weight queue. Furthermore, we also visualize the learned weight distributions, t-SNE embeddings, and measure the domain divergence and error set to show the good invariant learning capability of Unit-DRO. Empirical results show that the proposed Unit-DRO can effectively retrieve valuable samples or subgroups without demographics. 



In contrast to the word "Group" in Group-DROSagawa et al. (2019) where it assigns weights for domains, the word "Unit" in our proposed Unit-DRO assigns weights for samples.



Figure 1: An illustration of different person re-identification settings. (a) Supervised person ReID. (b) CD-ReID and UDA-ReID. (c) DG-ReID. (d) DGWD-ReID.

. Generalizable methods are recently proposed to learn invariant representations that can generalize to unseen Song et al. (2019); Choi et al. (2021); Zhang et al. (2021b).

