Enhancing Monte Carlo denoising with pixel-level guidance of exploiting auxiliary features

Abstract

The utilization of auxiliary features such as geometric buffers G-buffers and path descriptors P-buffers has greatly enhanced the denoising process in Monte Carlo (MC) techniques. However, recent methods let the neural network to implicitly learn how to exploit these auxiliary features, which may result in suboptimal utilization of each type. To address this issue, we propose a denoising framework that incorporates explicit pixel-wise guidance for leveraging auxiliary features. Our approach involves training two separate denoisers, each trained with a specific auxiliary feature G-buffers or P-buffers. We then employ an ensembling network to generate per-pixel ensembling weight maps, which serve as guidance for determining the dominant auxiliary feature for reconstructing each individual pixel. These weight maps are used to combine the outputs of the two denoisers. Additionally, we propagate the pixel-wise guidance to the denoisers by jointly training them with the ensembling network, encouraging the denoisers to focus on regions where G-buffers or P-buffers are more relevant for denoising. Our experimental results demonstrate significant improvement in denoising performance compared to a baseline model that utilizes both G-buffers and P-buffers.