In this paper, we describe a technique for the computation of augmented surface patches from stereo images. The stereo vision data is represented as patchlets, which are planar surface elements that have a position, surface normal, size, and confidence measures on the position and normal direction. To estimate the patchlet parameters, we use probabilistic image alignment. Whereas the patchlet generation method by Murray assumes the error on the disparity values to be equal for all pixels, we take the individual intensity patterns into account as well. This way, we obtain better estimates of the confidence measures in the patchlets. We describe the patchlet formation from the disparity and intensity values. To evaluate the quality of the results, we use real and synthetic images as they might be encountered in a typical mobile robot application. For a direct comparison, the images are processed both using Murray's and our approach.