Domain-Aware Unsupervised Cross-dataset Person Re-identification

We focus on the person re-identification (re-id) problem of matching people across non-overlapping camera views. While most existing works rely on the abundance of labeled exemplars, we consider a more difficult unsupervised scenario, where no labeled exemplar is provided. One solution for unsupervised re-id that attracts much attention in the recent researches is cross-dataset transfer learning. It utilizes knowledge from multiple source datasets from different domains to enhance the unsupervised learning performance on the target domain. In previous works, much effect is taken on extraction of the generic and robust common appearances representations across domains. However, we observe that there also particular appearances in different domains. Simply ignoring these domain-unique appearances will misleading the matching schema in re-id application. Few unsupervised cross-dataset algorithms are proposed to learn the common appearances across multiple domains, even less of them consider the domain-unique representations. In this paper, we propose a novel domain-aware representation learning algorithm for unsupervised cross-dataset person re-id problem. The proposed algorithm not only learns a common appearances across-datasets but also captures the domain-unique appearances on the target dataset via minimization of the overlapped signal supports across different domains. Extensive experimental studies on benchmark datasets show superior performances of our algorithm over state-of-the-art algorithms. Sample analysis on selected samples also verifies the ability of diversity learning of our algorithm.