Generalizing Semantic Part Detectors Across Domains

The recent success of deep learning methods is partially due to large quantities of annotated data for increasingly big variety of categories. However, indefinitely acquiring large amounts of annotations is not a sustainable process, and one can wonder if there exists a volume of annotations beyond which a task can be considered as solved or at least saturated. In this work, we study this crucial question for the task of detecting semantic parts which are often seen as a natural way to share knowledge between categories. To this end, on a large dataset of 15,000 images from 100 different animal classes annotated with semantic parts, we consider the two following research questions: (i) are semantic parts really visually shareable between classes? and (ii) how many annotations are required to learn a model that generalizes well enough to unseen categories? To answer these questions we thoroughly test active learning and DA techniques, and we study their generalization properties to parts from unseen classes when they are learned from a limited number of domains and example images. One of our conclusions is that, for a majority of the domains, part annotations transfer well, and that, performance of the semantic part detection task on this dataset reaches 98% of the accuracy of the fully annotated scenario by providing only a few thousand examples.