A Novel Segmentation Based Depth Map Up-Sampling

Fast, High-Quality Hierarchical Depth-Map Super-Resolution

“…Middlebury 2014 contains 33 datasets obtained using the technique of Scharstein et al [34] Implementation: Our system was built using Matlab on a PC with Core i7 2.9 GHz CPU and 32 GB RAM. In terms of datasets preprocessing, the high-resolution depth maps for these techniques are first inpainted (using SBU method) to remove the occlusions as shown in Figure 1, as some upsampling methods like JGU are only applicable to low-resolution depth maps with no occlusions [32]. Then, to generate the low-resolution depth, we downsample the inpainted image in 2×, 4×, 8×, 16×, respectively.…”

“…Thus we consider BPR to be particularly important in our evaluation of depth maps. We also use MSE_disc and BPR_disc to evaluate the performance near discontinuities [32].…”

“…Filtering-Based Depth-Map Super-Resolution. In most filteringbased methods, missing pixels are interpolated but existing depth pixels are left unchanged [21,24,25,32]. Joint bilateral upsampling (JBU) [21] is based on a joint bilateral filter [8,31], which is influenced by both color difference and spatial distances on depth map (e.g.…”

“…Joint trilateral filter based upsampling [25] first upsamples the low-resolution depth map with a simple interpolation method before refining the boundaries of the intermediate result in an outside-in order. Segmentation-based upsampling (SBU) [32] first converts the depth super-resolution problem into an RGB image segmentation problem. Based on this, a joint trilateral filter is constructed to locally interpolate the low-resolution depth.…”

“…Nevertheless, BPR and BPR_disc go extremely high. The SBU method [32] solves the above problems to a large extent based on the idea of segmentation. However, in boundary regions with rich textures, the segmentation result can be inaccurate, which may further degrade the final super-resolution result, as Figure 7(g) and Figure 8(g) show.…”