Monte Carlo Denoising with a Sparse Auxiliary Feature Encoder

Fast Denoising Monte Carlo path tracing is very desirable. Existing learning-based real-time methods concatenate auxiliary buffers (i.e., albedo, normal, and depth) with noisy colors as input. Such structures cannot effectively extract rich information from auxiliary buffers, however. In this work, we facilitate the U-shape kernel-prediction network with a sparse auxiliary feature encoder. Sparse convolutions can focus solely on regions whose inputs have changed and reuse the history features in other regions. With sparse convolutions, the computational complexity of the auxiliary feature encoder is reduced by 50-70% without apparent performance drops.