Color-guided depth refinement based on edge alignment

Depth maps captured by consumer-level depth cameras such as Kinect usually suffer from the problem of corrupted edges and missing depth values. In this paper, an effective approach with the support of guided color images is proposed to tackle this problem. Firstly, an effective two-pass alignment algorithm is used to reliably align the depth edges with color image edges. Then, a new depth map with refined edges is generated based on interpolated drift vectors. Finally, a constrained maximal bilateral filter is proposed to fill the holes. Compared with existing methods, our approach can better refine the depth edges and avoid blurred depths in areas of depth discontinuities, as demonstrated by experiments on real Kinect data.