Abstract
We present a simple algorithm for differentiable rendering of surfacesrepresented by Signed Distance Fields (SDF), which makes it easy to integraterendering into gradient-based optimization pipelines. To tacklevisibility-related derivatives that make rendering non-differentiable, existingphysically based differentiable rendering methods often rely on elaborateguiding data structures or reparameterization with a global impact on variance.In this article, we investigate an alternative that embraces nonzero bias inexchange for low variance and architectural simplicity. Our method expands thelower-dimensional boundary integral into a thin band that is easy to samplewhen the underlying surface is represented by an SDF. We demonstrate theperformance and robustness of our formulation in end-to-end inverse renderingtasks, where it obtains results that are competitive with or superior toexisting work.