Adaptive Rendering with Linear Predictions

摘要

We propose a new adaptive rendering algorithm that enhances thernperformance of Monte Carlo ray tracing by reducing the noise, i.e.,rnvariance, while preserving a variety of high-frequency edges in renderedrnimages through a novel prediction based reconstruction. Tornachieve our goal, we iteratively build multiple, but sparse linearrnmodels. Each linear model has its prediction window, where thernlinear model predicts the unknown ground truth image that can berngenerated with an infinite number of samples. Our method recursivelyrnestimates prediction errors introduced by linear predictionsrnperformed with different prediction windows, and selects an optimalrnprediction window minimizing the error for each linear model.rnSince each linear model predicts multiple pixels within its optimalrnprediction interval, we can construct our linear models only at arnsparse set of pixels in the image screen. Predicting multiple pixelsrnwith a single linear model poses technical challenges, related to derivingrnerror analysis for regions rather than pixels, and has not beenrnaddressed in the field. We address these technical challenges, andrnour method with robust error analysis leads to a drastically reducedrnreconstruction time even with higher rendering quality, comparedrnto state-of-the-art adaptive methods. We have demonstrated thatrnour method outperforms previous methods numerically and visuallyrnwith high performance ray tracing kernels such as OptiX andrnEmbree.