Gradient-Domain Vertex Connection and Merging

Abstract

Recently, gradient-domain rendering techniques have shown great promise in reducing Monte Carlo noise and improving overall rendering efficiency. However, all existing gradient-domain methods are built exclusively on top of Monte Carlo integration or density estimation. While these methods can be effective, combining Monte Carlo integration and density estimation has been shown (in the primal domain) to more robustly handle a wider variety of light paths from arbitrarily complex scenes. We present gradient-domain vertex connection and merging (G-VCM), a new gradient-domain technique motivated by primal domain VCM. Our method enables robust gradient sampling in the presence of complex transport, such as specular-diffuse-specular paths, while retaining the denoising power and fast convergence of gradient-domain bidirectional path tracing. We show that G-VCM is able to handle a variety of scenes that exhibit slow convergence when rendered with previous gradient-domain methods.