Structure-Texture Decomposition of RGB-D Images

In this paper, we study the problem of separating texture from structure in RGB-D images. Our structure preserving image smoothing operator is based on the region covariance smoothing (RCS) method in [16] that we present a number of modifications to this framework to make it depth-aware and increase its effectiveness. In particular, we propose to incorporate three geometric depth features, namely height above ground, angle with gravity and horizontal disparity to the pool of image features used in that study. We also suggest to use a new kernel function based on KL-divergence between the distributions of extracted features. We demonstrate our approach on challenges images from NYU-Depth v2 Dataset [24], achieving more accurate decompositions than the state-of-the-art approaches which do not utilize any depth information.