Depth map estimation: a region-based approach
Abstract
This paper presents an approach to the estimation of depth from stereo images which exploits correspondences of image segments. A key motivation is to bypass pixel correspondences which are ambiguous in texture-less regions. One of the main difficulty though is to come up to an equivalent partitioning of both stereo images that facilitates region matching. The reference image which the depth is estimated for is first segmented using the watershed algorithm. Regions are then transferred to the other image of the stereo pair according to a correlation analysis of region contours. This transfer yields regional disparities which are significant as long as the contours for a given region are not due to an occlusion. Based on that observation, we propose an algorithm that detects occlusion contours along a region and that rectifies disparity accordingly.
Origin : Files produced by the author(s)
Loading...