Image sampling for localization using entropy

This paper introduces a robust adaptive patches sampling technique. The method does not rely on the use of keypoints to extract local information but all information contained in images. It performs an optimal multilayer quadtree decomposition of images driven by the quantity and homogeneity of information. Extracted patches will be of different sizes according to the covered zones in the image and the information they contain. Experimental results carried out in localization, including different cases of corrupted images, and image topology. Finally to illustrate the technique possibilities, preliminary results in object recognition are shown.