Disparity map estimation with neural network

This work aims at defining a new approach for a dense disparity map computing based on the neural networks from a pair of stereo images. Our approach has been divided into two main tasks. The first one deals with computing the initial disparity map using a neuronal method (BP). Whereas the second one presents a simple method to refine the initial disparity map using neural refinement so that an accurate result can be acquired. In the literature, the matching score is based only on the pixel intensities. We introduce in this work two additional features: the gradient magnitude and orientation of the gradient vector of pixels which gives a true degree of similarity between pixels. Experimental results on real data sets were conducted for evaluating the proposed method.