Depth Map Super-Resolution by Deep Multi-Scale Guidance: Supplementary material

We intend to show that the optimal filter size of backwards convolution (or deconvolution (deconv)) for upsampling is closely related to the upscaling factor s. For conciseness, we consider a single-scale network (SS-Net(ord)) trained in an ordinary domain for upsampling a LR depth map with an upscaling factor s = 4. Figure 1 shows an overview of SS-Net(ord). Specifically, the first and third layers perform convolution, whereas the second layer performs backwards strided convolution. Activation function PReLU is used in SS-Net(ord) except the last layer. We set the network parameters: n1 = 64, n2 = 32, n3 = 1 and f1 = f3 = 5. We evaluate the super-resolving performance of SS-Net(ord) by using different deconv filter sizes f2× f2. Figure 2 shows the convergence curves using f2 ∈ (3, 9, 11). It can be shown that upsampling accuracy increases with f2 until it reaches 2s+1 i.e. f2 = 9. In a compromise between computation efficiency and upsampling performance, we choose deconv filter size to (2s+ 1)× (2s+ 1).