Multi-source adaptive meta-learning framework for domain generalization person re-identification.

Multi-source domain generalization for person re-identification, named Domain Generalizable Person Re-Identification (DG-ReID), refers to training on multiple source domains and testing on unseen target domains. However, the performance of the current methods is rather unsatisfactory. In this paper, we propose the Multi-Source Adaptive Meta-Learning (MAM) framework to push the limit of DG-ReID. First, MAM is built on the novel Multi-Source Adaptive Feature Refinement (MAFR) block, which simultaneously acquires domain-generalized features and domain-unique features on different domains. MAFR consists of three key components: the fine-grained feature extraction (FGFE) branch, the refined feature supplementation (RFS) branch, and the feature fusion (FF) module. Through these blocks, the interrelationships between different domains are explored, facilitating the extraction of features that are both generalized and discriminative. To capture domain-adaptive information, MAFR incorporates a meta-weight generator to effectively capture domain-adaptive information. Second, MAM leverages the principles of meta-learning during training to acquire more powerful generalization features. A total of three novel loss functions are proposed to train the proposed MAM including the clustering loss, the unknown domain identification loss, and the non-parametric identification loss. Last, we further design a novel meta-feature sampling method to diversify features used in the meta-testing stage for better generalization capability. The effectiveness of the proposed method is extensively evaluated on multiple benchmarks, and the results provide comprehensive validation. For instance, our MAM framework presents a promising solution to the challenge of domain generalization in person re-identification. Specifically, under evaluation protocol-2, the MAM framework exhibits improvements of 1.0% and 0.9% in mAP and rank-1 metrics, respectively, compared to the previous state-of-the-art method. [ABSTRACT FROM AUTHOR]