Efficient Dense Depth Estimation from Dense Multiperspective Panoramas

In this paper we study how to compute a dense depth map with panoramic field of view (e.g., 360 degrees) from multiperspective panoramas. A dense sequence of multiperspective panoramas is used for better accuracy and reduced ambiguity by taking advantage of significant data redundancy. To speed up the reconstruction, we derive an approximate epipolar plane image that is associated with the planar sweeping camera setup, and use one-dimensional window for efficient matching. To address the aperture problem introduced by one-dimensional window matching, we keep a set of possible depth candidates from matching scores. These candidates are then passed to a novel two-pass tensor voting scheme to select the optimal depth. By propagating the continuity and uniqueness constraints noniteratively in the voting process, our method produces highquality reconstruction results even when significant occlusion is present. Experiments on challenging synthetic and real scenes demonstrate the effectiveness and efficacy of our method.