Learning Generic Local Shape Properties for Adaptive Super-Sampling

We propose a novel encoder/decoder-based neural network architecture that learns view-dependent shape and appearance of geometry represented by voxel representations. Since the network is trained on local geometry patches, it generalizes to arbitrary models. A geometry model is first encoded into a sparse voxel octree of features learned by a network, and this model representation can then be decoded by another network in-turn for the intended task. We utilize the network for adaptive super-sampling in ray-tracing, to predict super-sampling patterns when seeing coarse-scale geometry. We discuss and evaluate the proposed network design, and demonstrate that the decoder network is compact and can be integrated seamlessly into on-chip ray-tracing kernels. We compare the results to previous screen-space super-sampling strategies as well as non-network-based world-space approaches.