Specular Lobe Aware Upsampling Based on Spherical GaussiansSupplemental Material)

Figure 1 shows the comparison of a general normal based weighting function and the proposed weighting function. They are rendered with voxel cone tracing [Crassin et al. 2011] and spatio-temporal upsampling [Herzog et al. 2010]. The scene has a dynamic object (a galloping horse). In this experiment, the normal based weighting function has a Gaussian distribution [Yang et al. 2008] and its variance parameter is 0.005. Other weighting functions described in [Herzog et al. 2010] such as a spatial weighting function, a depth based weighting function, and a dynamic temporal weighting function using temporal gradients are also used for all images. Unlike original spatio-temporal upsampling, we additionally apply our weighting function or the normal based weighting function to the temporal axis. This approach efficiently reduces undesirable temporal blurring artifacts for dynamic objects.