Gradient Outlier Removal for Gradient-domain Path Tracing

Abstract

We propose a new algorithm of outlier removal for Gradient-domain Path Tracing (G-PT) which computes the gradients of an image in addition to its colors. Our main purpose is to reject gradient outliers that are estimated as having higher errors than others for robust reconstruction followed by rendering quality improvements. We present a Least Trimmed Squares (LTS) method adjusted for our gradient outlier problem, which trims partial gradients considered as outliers so that only the optimal subset of gradients is used to reconstruct the final image. For boosting efficiency, we design a new approach of modeling our problem into a 2D grid and guaranteeing connectivity between each pixel through gradients, so that it prevents an isolated pixel from being left out of reconstruction. Finally, we find the optimal number of gradients to be used in our reconstruction using LTS by estimating the reconstruction error. We demonstrate that our method with effectively removing gradient outliers contributes to improvements visually and numerically for G-PT, which is previously using all the gradients for screened Poisson reconstruction.