Abstract

Feature-based object matching is a fundamental problem in computer vision. In this talk, we present a new first-order object (inlier features) matching technique called ROML (Robust Object Matching using Low-rank constraint). Given a set of images with extracted inlier and outlier features, ROML aims to simultaneously identify the inlier features from each image, and establish their consistent correspondences across the image set. This is a challenging combinatorial problem. To achieve the goal, ROML leverages the underlying data low-rank property to simultaneously optimize a partial permutation matrix (PPM) for each image, and feature correspondences are established by the obtained PPMs. Extensive experiments on rigid/non-rigid object matching, matching instances of a common object category, and common object localization demonstrate ROML’s efficacy for feature-based object matching.