IPAB workshop - 26/01/2017

Title: Neural Network Ambient Occlusion

Abstract: Screen Space Ambient Occlusion is a technique for producing an approximation of the soft shadows of a scene from just the depth map.

This technique is commonly used in game rendering as it is simple, fast, and independent of geometric complexity. Since the depth map does not provide complete information about the scene, previous algorithms approach this task in an ad-hoc way. In this talk I present a new algorithm called Neural Network Ambient Occlusion which approaches this problem using machine learning. Using a large database of ground truth renders made offline we attempt to learn the Screen Space Ambient Occlusion effect using a specially designed neural network which, once trained, is faster and more accurate than previous methods.