Adaptive rendering based on a weighted mixed-order estimator

In this paper, we propose a novel adaptive rendering method to robustly handle noise artifacts and outliers of Monte Carlo ray tracing by combining the Nadaraya-Watson and robust local linear estimators while efficiently preserving fine details. Our method first constructs a sparse robust local linear estimator in feature space (normal,texture,etc.), while also removing spike noise. Then, we utilize the Nadaraya-Watson estimator to filter the outlier-free image. We generate the final image by interpolating the values of each estimator at each pixel with weights that are inversely proportional to the estimated mean squared errors. Lastly, we distribute additional samples to the regions with higher estimated mean squared errors if sampling budget remains. We have demonstrated that our estimator outperforms previous methods visually and numerically.