Articulated 3-D modelling in a wide-baseline disparity space

Image-based novel-view synthesis requires dense correspondences between the original views to produce a high quality synthetic view. In a wide-baseline stereo setup, dense correspondences are difficult to achieve due to the significant change in viewpoint giving rise to a number of problems. To improve their quality, the original, incomplete disparity maps are usually interpolated to fill in the missing regions. When the data is very sparse, such as in the case of the wide-baseline stereo, interpolation alone is not enough. Instead, a 3-D model of the scene can be used to fill in the missing regions more reliably, using a-priori knowledge. However, the 3-D model can be used more efficiently and accurately in disparity space, where the disparity data originates from. In this paper we present and compare the two techniques. We show that, in comparison with the 3-D approach, the disparity space approach offers a computationally less expensive and potentially more accurate solution.