2016-ICCV-Domain Adaptation for Object Recognition：An Unsupervised Approach.pdfVIP

Domain Adaptation for Object Recognition: An Unsupervised Approach∗ Raghuraman Gopalan, Ruonan Li, and Rama Chellappa Center for Automation Research, University of Maryland, College Park, MD 20742 USA {raghuram,liruonan,rama}@ Abstract ferent. On the other hand, DA pertains to the case where ˜ ˜ P (Y |X) ≈ P (Y |X), but P (X) significantly varies from ˜ Adapting the classifier trained on a source domain to P (X). This specific scenario occurs very naturally in un- recognize instances from a new target domain is an impor- constrained object recognition settings, where the domain tant problem that is receiving recent attention. In this pa- shift can be due to change in pose, lighting, blur, and reso- per, we present one of the first studies on unsupervised do- lution, among others. main adaptation in the context of object recognition, where Understanding the effects of domain change has received we have labeled data only from the source domain (and substantial attention from the natural language processing therefore do not have correspondences between object cat- community over the last few years (e.g. [4, 8, 16]). Al- egories across domains). Motivated by incremental learn- though many fundamenta