3-D Volumetric Shape Abstraction from a Single 2-D Image

We present a novel approach to recovering the qualitative 3-D part structure from a single 2-D image. We do not assume any knowledge of the objects contained in the scene, but rather assume that they are composed from a user-defined vocabulary of qualitative 3-D volumetric part categories input to the system. Given a set of 2-D part hypotheses recovered from an image, representing projections of the surfaces of the 3-D part categories, our method simultaneously selects and groups subsets of the 2-D part hypotheses into 3-D part "views", from which the shape and pose parameters of the volumetric parts are recovered. The resulting 3-D parts and their relations offer the potential for a domain-independent, viewpoint-invariant shape indexing mechanism that can help manage the complexity of recognizing an object from a large database.