Depth interpolation via segmentation defined from smooth junctions of color geodesic neighborhoods

This paper describes an iterative procedure of accurate depth interpolation for a low-resolution depth image using a high-resolution color image. In our procedure, depth interpolation is gradually achieved with generating smooth junctions of similar color regions. These junctions give the segmentation of the interpolated region and we use this information to interpolate depth value on each segment, and to promote the generation of junctions additionally. In experiments with images of the Middlebury stereo datasets, our procedure interpolates each image at a high rate, and achieves the low error compared with previous methods.