Uncertainty-Aware RGBD Image Segmentation

We propose a graph-based RGBD image segmentation method that considers both depth and color information. Color and depth information are complementary to each other. However, compared with the RGB channels, the depth channel of an image has more noises and uncertainties that have negative effects to accurate segmentation. To partially solve this problem, we model the depth uncertainties of an image as a function of the distances and angles between the RGBD sensor and the observed surfaces. Then, the uncertainty model is applied to RGBD image segmentation in which the RGB and depth cues are combined according to the uncertainties of the depth measurements. The experimental results show that our method improves the segmentation accuracy.