2D Parameter Filtering for Noisy Gradient in Differentiable Rendering

Abstract

Differentiable rendering is a method that can estimate the gradient for scene parameters from the rendering process, and this gradient can used to solve the inverse rendering problem through a gradient-based optimization method. In particular, the physically-based differentiable rendering is a method that can estimate gradients that can estimate parameter values more accurately than other methods, even in images with complex light transport effects. However, physically-based differentiable rendering introduces high variance into the gradient when the number of rendering samples is insufficient, which makes the convergence of the gradient-based optimization method challenging. To address this problem, we propose a 2D parameter filtering method using spatial filters in an inverse rendering pipeline using physically-based differentiable rendering. Our method shows better optimization than existing methods by reducing the gradient variance that appears as noise as the back-propagated gradient passes through a spatial filter.