Dense Disparity Maps Respecting Occlusions and Object Separation Using Partial Differential Equations

摘要

In this work, we present substantial enhancements to solve the stereo correspondence problem using a minimization and regulariza-tion formulation with a partial differential equations approach. For the first time it allows to respect occlusions and separation of objects. We introduce a boundary condition that makes it possible to estimate disparities of arbitrarily shaped regions and thus to explicitely handle occlusions. After calculating a dense disparity map, we detect occlusions and object boundaries, cut the disparity map at these boundaries, and minimize the energy functional once more resulting in more accurate estimates. We show that we can achieve a speed up of a factor of four by rectified image pairs and a correlation based algorithm to calculate an initial estimate. In addition, a priori knowledge provided as region of interest or location of object boundaries further improves both the speed and the quality of the estimation. The results demonstrate that the quality is greatly improved with the proposed approach.