Interpolating Novel Views from Image Sequences by Probabilistic Depth Carving

We describe a novel approach to view interpolation from image sequences based on probabilistic depth carving. This builds a multivalued representation of depth for novel views consisting of likelihoods of depth samples corresponding to either opaque or free space points. The likelihoods are obtained from iterative probabilistic combination of local disparity estimates about a subset of reference frames. This avoids the difficult problem of correspondence matching across distant views and leads to an explicit representation of occlusion. Novel views are generated by combining pixel values from the reference frames based on estimates of surface points within the likelihood representation. Efficient implementation is achieved using a multiresolution framework. Results of experiments on real image sequences show that the technique is effective.