Object recognition using local affine frames on maximally stable extremal regions

Methods based on distinguished regions (transformation covariant detectable regions) have achieved considerable success in object recognition, retrieval and matching problems in both still images and videos. The chapter focuses on a method exploiting local coordinate systems (local affine frames) established on maximally stable extremal regions. We provide a taxonomy of affine-covariant constructions of local coordinate systems, prove their affine covariance and present algorithmic details on their computation. Exploiting processes proposed for computation of affine-invariant local frames of reference, tentative region-to-region correspondences are established. Object recognition is formulated as a problem of finding a maximal set of geometrically consistent matches. State of the art results are reported on standard, publicly available, object recognition tests (COIL-100, ZuBuD, FOCUS). Change of scale, illumination conditions, out-of-plane rotation, occlusion, locally anisotropic scale change and 3D translation of the viewpoint are all present in the test problems.