An Efficient Dense Depth Map Estimation Algorithm Using Direct Stereo Matching for Ultra-Wide-Angle Images

We present an efficient dense depth map estimation algorithm using patch-based direct stereo matching for ultra-wide-angle images. Our algorithm takes account of the fact that the neighboring pixels inside a local patch are likely to lie on the same plane. Our algorithm propagates the “good” initial guesses to the neighboring pixels by spatial propagation, followed by a random refinement process. Therefore, this allows finding precise depth value for each point in an infinite space using a random search strategy. Our algorithm can be used to perform 3D reconstruction using the dense depth maps directly generated from ultra-wide-angle images, especially from stereo camera pairs.