Automatic multiview coarse registration of range images for 3D modeling

Automatic registration of different overlapping range images (views) of an object is performed by matching their features. This process, known as correspondence identification, registers the views coarsely and is followed by fine registration. Existing correspondence techniques are limited to a pair of views at a time. In case there are more than two views of an object in random order, these techniques must perform an exhaustive search for correspondences between all possible view pairs. In this paper, we present an efficient automatic multiview surface matching algorithm for 3D modeling which simultaneously matches a single view with multiple views. Our approach represents local surface patches of each view with multiple tensors. Tensors are indexed by a 4D hash table. A voting approach is used with the help of the hash table to simultaneously identify potential corresponding tensors of different views. Correspondences are verified and used to construct a spanning tree graph with nodes representing the views and arcs representing the rigid transformation that aligns two views. This graph is used to register all the views. Our results show that our algorithm is efficient and robust to noise.