Object co-detection

Abstract

In this paper we introduce a new problem which we call object co-detection. Given a set of images with objects observed from two or multiple images, the goal of co-detection is to detect the objects, establish the identity of individual object instance, as well as estimate the viewpoint transformation of corresponding object instances. In designing a co-detector, we follow the intuition that an object has consistent appearance when observed from the same or different viewpoints. By modeling an object using state-of-the-art part-based representations such as [1,2], we measure appearance consistency between objects by comparing part appearance and geometry across images. This allows to effectively account for object self-occlusions and viewpoint transformations. Extensive experimental evaluation indicates that our co-detector obtains more accurate detection results than if objects were to be detected from each image individually. Moreover, we demonstrate the relevance of our co-detection scheme to other recognition problems such as single instance object recognition, wide-baseline matching, and image query. © 2012 Springer-Verlag.