Disambiguating the recognition of 3D objects

Abstract

We propose novel algorithms for the detection, segmentation, recognition, and pose estimation of three-dimensional objects. Our approach initially infers geometric primitives to describe the set of 3D objects. A hierarchical structure is constructed to organize the objects in terms of shared primitives and relations between different primitives in the same object. This structure is shown to disambiguate the object models and to improve recognition rates. The primitives are obtained through our new Invariant Hough Transform. This algorithm uses geometric invariants to compute relations for subsets of points in a specific object. Each relation is stored in a hash table according to the invariant value. The hash table is used to find potential corresponding points between objects. With point matches, pose estimation is achieved by building a probability distribution of transformations. We evaluate our methods with experiments using synthetic and real 3D objects.