Boosting local matches with convolutional co-segmentation

Matching corresponding local patches between images is a fundamental building block in many computer-vision algorithms. Most matching methods are composed of two main stages: feature extraction, typically done independently on each image, and feature matching which is done on processed representations. This strategy tends to create large amounts of matches, typically describing small, highly-textured regions within each image. In many cases, large portions of the corresponding images have a simple geometric relationship. We exploit this fact and reformulate the matching procedure to an estimation stage, where we extract large domains roughly related by local transformations, and a convolutional Co-Segmentation stage, for densely detecting accurate matches in every domain. Consequently, we represent the geometrical relationship between images with a concise list of accurately co-segmented domains, preserving the geometrical flexibility stemmed from local analysis. We show how the proposed co-segmentation improves the matching coverage to accurately include many low-textured domains.