Class-driven color transformation for semantic labeling

Abstract

We propose a novel class-driven color transformation aimed at semantic labeling. In contrast with other approaches elsewhere in the literature, our approach is a supervised one employing class information to learn a color transformation. Our method maps image color to a target space with maximum pairwise distances between classes and minimum scattering within each of them. To compute the color transformation, we pose the problem in terms of a composition of two mappings. The first mapping employs a pairwise discriminant cost function minimized through a steepest descent optimization to map the image color data onto a space spanned by the class set. It targets better separability between distinct classes as well as less scattering within each individual class. The second mapping corresponds to subspace projection of this class data to a target space with same dimensionality of image color data. To preserve distances attained by the first of the mappings, this subspace projection is effected making use of metric multi-dimensional scaling. We report our experiments on MSRC-21 and SBD datasets, where our method consistently improves overall and average performances of well-known publicly available TextonBoost and DARWIN multiclass segmentation frameworks at a negligible computational cost. These results confirms our contribution towards reflection of higher distinction in color space by imposing better separability in a novel representation which is learned from class information of the dataset under consideration.