Disparity estimation in stereoscopic vision by simulated annealing

This paper presents a correspondence algorithm for stereo vision based on an integrated model that includes several units corresponding to different stages: features extraction, minimization of an energy function using simulated annealing, multiresolution and interpolation. Firstly the original images are scaled down to considerably reduce their size. From the reduced images a disparity map is obtained, which is used as the basis to develop the complete process. For this reason an energy function is built and minimized using a multiresolution scheme. The energy function integrates features such as grey level, non. parametric transforms, edges, smoothness and uniqueness. The obtained disparity for every resolution is interpolated to work with the following resolution. Our model produces a dense disparity map. The algorithm has been tested with different kinds of real images to show its flexibility.