Depth Restoration with Normal-Guided Multiresolution Superpixel

Jinghui Qian, Jie Guo, Jingui Pan

Abstract: In this paper, we propose a depth restoration method using a novel superpixel technique. Guided by a normal map reconstructed from the raw depth data, this technique over-segments RGB-D images into many small regions where their depth is assumed to be smooth. As the raw depth data is incomplete, we further introduce a depth confidence map to identify the regions which are more reliable. With the produced superpixels, we can restore the incomplete depth map using a per-superpixel linear regression. A multiresolution superpixel strategy is employed when some superpixels do not contain enough valid data. Experiments show that the proposed depth restoration method can effectively fill the wide gaps along depth discontinuities without blurring the object boundaries and the depth discontinuities.

External IDs:dblp:conf/icmcs/Qian0P18