Generalizable person re-identification with part-based multi-scale network

Supervised person re-identification (Re-ID) has advanced significantly, but it suffers from the performance drop when the pretrained models are directly deployed to an unseen domain. Meanwhile, domain adaptation methods are widely investigated to decrease the performance degradation caused by domain gaps. However, it still requires training with unlabeled target-domain data and iteratively updating models. In this work, we proposed a generalizable person Re-ID framework named Part-based Multi-scale Network (PMN), which was trained on source domain(s) once and can be directly exploited to target domains with stable performance. To this end, we leveraged a part-based architecture which uniformly partitions feature maps into several horizontal stripes. The stripe features contain fine-grained information of human parts and therefore benefit learning discriminative features. The Scale Adjusting Module (SAM) is also designed to regulate the style differences appearing in lower-level feature maps and helps incorporation of features from different levels. When we integrated the style-adjusted features and fine-grained local features into our improved backbone, the proposed framework becomes generalized to variation of image styles and backgrounds from different datasets. Extensive experiments show the superiority of the proposed PMN over state-of-the-art generalizable methods on multiple popular Re-ID benchmarks with cross-domain setting. Furthermore, we also demonstrate the advantage of using our framework as a backbone for domain adaptation methods.