Abstract:
Solving stereo vision by minimization and regularization methods was recently introduced by Robert and Deriche, based on techniques used in restauration. This method calculates depth directly, without finding disparities. It is based on an iterative multiresolution algorithm where the depth of each pixel in the reference image is calculated directly by minimizing an energy function. Depth discontinuities are preserved using an appropiate regularization term. In this technical report we extend this method, using the neighbourhood of each pixel for computing the disimilarity between two pixels in two images. The new approach pretends to give more consistency to each pixel processing in the iterative procedure in order to obtain better results. Here, a theoretical study is presented.