Unsupervised segmentation of RGB-D images

Abstract

While unsupervised segmentation of RGB images has never led to results comparable to supervised segmentation methods, a surprising message of this paper is that unsupervised image segmentation of RGB-D images yields comparable results to supervised segmentation. We propose an unsupervised segmentation algorithm that is carefully crafted to balance the contribution of color and depth features in RGBD images. The segmentation problem is then formulated as solving the Maximum Weight Independence Set (MWIS) problem. Given superpixels obtained from different layers of a hierarchical segmentation, the saliency of each superpixel is estimated based on balanced combination of features originating from depth, gray level intensity, and texture information. We want to stress four advantages of our method: (1) Its output is a single scale segmentation into meaningful segments of a RGB-D image; (2) The output segmentation contains large as well as small segments correctly representing the objects located in a given scene; (3) Our method does not need any prior knowledge from ground truth images, as is the case for every supervised image segmentation; (4) The computational time is much less than supervised methods. The experimental results show that our unsupervised segmentation method yields comparable results to the recently proposed, supervised segmentation methods [1,2] on challenging NYU Depth dataset v2.