Multiscale TILT Feature Detection with Application to Geometric Image Segmentation

Motivated by the theory of low-rank matrix representation, a new type of invariant image feature, called transform-invariant low-rank texture (TILT), has been recently proposed. However, the application of TILT features in computer vision has been severely limited by two major problems. First, TILT feature representation is based on the assumption that the given image contains only one dominant low-rank region, which typically does not hold in natural images. Second, when multiple low-rank regions are present, the existing TILT detection methods either randomly sample the image or apply to fixed grid coordinates, both of which cannot guarantee good recovery of salient low-rank image features. In this paper, we propose a novel algorithm to address these two important issues. First, utilizing super pixels and the concept of canonical rank derived from TILT, we introduce a method to segment natural images into a geometric layer and a non-geometric layer. Second, we apply a Markov random field model to a multiscale low-rank representation of the image geometric layer, and obtain an effective algorithm to detect TILT features. Finally, we present an application of the multiscale TILT detection algorithm to the classical problem of building facade segmentation. Extensive experiments are conducted on the Pankrac building database to demonstrate the efficacy of the algorithms.