The dominant approach for learning local patch descriptors relies on small
image regions whose scale must be properly estimated a priori by a keypoint
detector. In other words, if two patches are not in correspondence, their
descriptors will not match. A strategy often used to alleviat