Given a pair of stereo images and the transformation existing between the corresponding camera coordinate systems, the depth of a scene point can be computed from its projections on both images. Despite the difficulties related to the matching of such projections across homogeneous regions and the occlusion phenomenon, state of the art methods have already produced accurate results on classical stereo datasets. This article proposes a new way of approaching depth estimation. Instead of searching for dense pixel correspondences, a gross estimation of the disparities is initially performed at the region level, resulting in a regional disparity map which highlights the principal depth layers of the image. The disparity map is then systematically refined by considering finer partitions of the image. To this end, the watershed of the image colour gradient is selected in order to compute the image partitions alongside a meaningful hierarchy. We show that the ability to be driven by labelled markers enables the watershed algorithm to generate a co-segmentation of both stereo images given the regional disparities, which constitutes the main contribution of this paper. This co-segmentation allows one to reliably compute the disparities of pixels along the region contours. Finally, the contour disparities are transferred to the concerned regions after a careful analysis of their occlusion state with respect to each adjacent region. Though approximate, we show that the proposed method yields regional disparity maps which are close enough to ground truths in the view of performing the desired refinements. We also expose the perspectives of this methodology with respect to challenging stereo imagery, i.e. which is affected by noise or which contains a considerable amount of homogeneous regions.
展开▼
