Accurate Real-Time Disparity Estimation with Variational Methods

Estimating the disparity field between two stereo images is a common task in computer vision, e.g., to determine a dense depth map. Variational methods currently are among the most accurate techniques for dense disparity map reconstruction. In this paper a multi-level adaptive technique is combined with a multigrid approach that allows the variational method to achieve real-time performance (on a CPU). The multi-level adaptive technique refines the grid only at peculiarities in the solution. Thereby it reduces the computational effort and ensures that the reconstruction quality is kept almost the same. Further, we introduce a technique that adapts the regularizer, used in the variational approach, dependend on the the current state of the optimization. This improves the reconstruction quality. Our real-time approach is evaluated on standard datasets and it is shown to perform better than other real-time disparity estimation approaches.