Efficient Surface Interpolation with Occlusion Detection

In this paper we present a novel dense matching algorithm that relies on sparse stereo data in order to build a dense disparity map. The algorithm uses a recursive updating scheme to estimate the dense stereo data using various interpolation techniques. The major problem of classical template matching techniques is their reliance on a fixed template shape and poor performance around untextured regions. In this paper we attempt to alleviate the problem of template matching techniques by using an adaptive window shape and also by avoiding searching in homogenous image regions that are difficult to match by templates. The outcome is an algorithm that performs at least ten times faster than template matching, and yet it achieves higher accuracy. Moreover, our algorithm preserves depth discontinuities and assigns disparities at occluded regions.