Depth Map Enhancement Using Adaptive Steering Kernel Regression Based on Distance Transform

Abstract

In this paper, we present a method to enhance noisy depth maps using adaptive steering kernel regression based on distance transform. Data-adaptive kernel regression filters are widely used for image denoising by considering spatial and photometric properties of pixel data. In order to reduce noise in depth maps more efficiently, we adaptively refine the steering kernel regression function according to local region structures, flat and textured areas. In this work, we first generate two distance transform maps from the depth map and its corresponding color image. Then, the steering kernel is modified by a newly-designed weighing function directly related to joint distance transform. The weighting function expands the steering kernel in flat areas and shrinks it in textured areas toward local edges in the depth map. Finally, we filter the noise in the depth map with the refined steering kernel regression function. Experimental results show that our method outperforms the competing methods in objective and subjective comparisons for depth map enhancement.