A minimization approach for 3D recovery in region-based stereo vision

Stereoscopic vision is the set of techniques that try to recover three-dimensional information from two or more views of a scene. The usual technique consists of computing correspondences between points in both images in order to be able to calculate their depth, thus obtaining their three-dimensional location. Usually, the correspondences are represented by a disparity map from which we can obtain the depth map of the reference image. An important problem to take into account while computing correspondences is the presence of depth discontinuities in the scene. Several techniques try to solve this problem during the matching process. We propose using regions as the matching primitive, given that they represent homogeneous areas limited by intensity discontinuities. We impose some constraint in the shape of depth inside the regions, while allowing depth discontinuities at the region boundaries. The used constraint is an important issue that is discussed. The use of regions has been already proposed usually as an initial stage of the correspondence problem. These techniques have to be able to handle segmentation errors, often by merging some adjacent regions in order to improve correspondence. In our approach, we propose considering the regions resulting from segmentation of a reference image and finding their correspondences without segmenting the other image. The basis of the method and a detailed explanation of the algorithm for region matching and occlusion detection are presented. The experimental results of the method are shown. Finally, we discuss the conclusions extracted from the experiments carried out so far