Instance interactive association graph convolutional network for domain adaptive person re-identification

Domain adaptive person re-identification (re-ID) is a challenging task due to the large domain divergency between different datasets and the complicated variations of the target domain. Style-transferring based methods mainly narrow the domain divergency with respect to a specific imaging factor, e.g., illumination. However, in consideration of the complex practical scenarios, domain adaptive re-ID demands comprehensive domain characteristic knowledge, i.e., seasons, illuminations, camera views, etc., to alleviate the domain divergency and intra-domain variations. This paper proposes a data-driven Instance Interactive Association Graph Convolutional Network (IIAGCN) to tackle the problems. Specifically, our IIAGCN method first constructs a cross-domain knowledge graph with the inter-domain nodes (target-source image pairs) and the intra-domain nodes (target-target image pairs). Then an Information Interactive Graph Convolutional (IIGC) layer is designed to extract the instance-level domain characteristic knowledge from the knowledge graph. With the learned knowledge, we can learn domain characteristic-aware and discriminative image representations for better domain adaptation. In addition, we introduce the memory bank component to store image features of the whole dataset, which enlarges the node diversity of the knowledge graph. Experiments on large-scale person re-ID datasets demonstrate the superiority of our method under the unsupervised re-ID setting.