Fundamentals of (Neural) Inverse Rendering

Andrea Tagliasacchi

Simon Fraser University, CAN

Abstract

The capture of 3D content from multiple images has been revolutionized by the availability of convenient to use auto-differentiation packages. This has enabled leaps in visual performance, as it moved the field from reconstructing texture-less surfaces given point clouds, to reconstructing photorealistic scenes from collections of images via volume rendering. In this talk I will describe the fundamentals of inverse rendering, the underlying representations that are commonly used, and identify some of the typical assumptions that these methods impose on the training data.