Image Clustering under Domain Shift

We address domain adaptation in the context of clustering where we are given a set of unlabeled data, coming from several domains, and the goal is to group data into different categories regardless of the domain they come from. This is a challenging problem since we do not have any supervision unlike most adaptation scenarios studied earlier, and is very relevant in practical industry applications where labeled data often comes at a premium especially while deploying services that do not have a comparable predecessor. Our philosophy in addressing this problem draws motivation from the concept of dirty paper coding, a communications technique where the signal being transmitted through a noisy channel is encoded with priors on the possible noise patterns to assist reliable decoding of the signal at the receiver. We focus on image applications in this paper, where we encode priors on possible image domain shift factors such as viewpoint, lighting, blur and appearance variations and utilize geometric adaptation mechanisms to perform clustering. We illustrate the utility of our approach on standard datasets involving objects and faces, by obtaining around 18% improvement on average over existing approaches.