Adaptive rendering based on robust principal component analysis

We propose an adaptive sampling and reconstruction method based on the robust principal component analysis (PCA) to denoise Monte Carlo renderings. Addressing spike noise is a challenging problem in adaptive rendering methods. We adopt the robust PCA as a pre-processing step to efficiently decompose spike noise from rendered image after the image space is sampled. Then we leverage patch-based propagation filter for feature prefiltering and apply the robust PCA to reduce dimensionality in high-dimensional feature space. After that, we estimate a per-pixel pilot bandwidth derived from kernel density estimation and construct the multivariate local linear estimator in the reduced feature space to estimate the value of each pixel. Finally, we distribute additional ray samples in the regions with higher estimated mean squared error if sampling budget remains. We demonstrate that our method makes significant improvement in terms of both numerical error and visual quality compared to the state-of-the-art.