Heterogeneous graph driven unsupervised domain adaptation of person re-identification

Abstract

How to incrementally optimize a pre-trained classifier in an unlabeled target domain is a core challenging problem of domain adaptation (DA) for many visual tasks, such as Person Re-identification (re-ID). Most of the existing methods optimize the model based on pseudo labels or similarity of instance pairs, but ignoring the diverse manifold structures of unlabeled instances in the whole dataset. In this paper, we address the importance of such structural information in domain adaptation, and propose a Heterogeneous Graph driven Optimization scheme, namely H-GO, for structure based unsupervised learning. In particular, H-GO builds a heterogeneous graph of unlabeled images to consider the heterogeneous properties of images from various cameras with varied visual styles. A heterogeneous affinity propagation method is further applied to explore the graph based affinity between the instances which share similar manifold structures. Finally, a heterogeneous affinity learning procedure is taken to optimize the visual models by using the graph based affinity of instances. Comprehensive experiments are conducted on three large-scale re-ID datasets, and the results demonstrate the flexibility and the superior performance of H-GO than state-of-the-art unsupervised domain adaptation algorithms.