Anchor-ReID: A Test Time Adaptation for Person Re-identification

Abstract

Person re-identification (ReID) is a challenging computer vision problem where the objective is to retrieve a person of interest from a gallery of images. Conventional person ReID methods struggle to generalize across different domains, leading to inferior cross-domain performance. To address this, domain generalization (DG) person re-id methods are proposed. The majority of DG studies focus on designing more robust models that are trained on source datasets in an offline setting with explicit components or training mechanisms for domain generalization. Moreover, these frameworks do not take into consideration that in the real world the environment is in continual change resulting in the degradation of the discriminative feature extraction. To tackle these problems, we propose a new problem formulation of Test-time adaption for ReID (TTA ReID). Our TTA ReID takes a pre-trained model that was not designed for domain generalization from source domain, and aims in inference to adapt it to the current target domain in an unsupervised setting. To this end, we propose Anchor ReID, a framework designed to adapt a pre-trained model without altering its network architecture and make it robust to domain shift. Comprehensive experiments on CUHK03, Duke-mtmc, and Market 1501 datasets demonstrate the benefits of the proposed approach. The proposed Anchor ReID framework can improve a pre-trained model (that was not designed for domain generalization) and achieves an absolute gain of mAP 10 in the CUHK03 dataset.