Learning to light and relight images

点击翻译按钮获取中文摘要

Augmented reality is the process of combining the virtual with the real. For the combination to look realistic, the virtual content must match the real: their orientation, lighting, and presence in the scene must all match the real-world conditions. Achieving highly realistic effects is possible (as demonstrated on a regular basis in blockbuster movies), but doing so requires tremendous amounts of manpower---indeed, all special effects nowadays require tens, if not hundreds, of skilled artists painstakingly working on each shot. In this proposal, we aim at taking a step towards automating the process of realistically inserting virtual objects into real photographs. Specifically, we address the problem of lighting estimation along the following two main axes. First, we will demonstrate that the full, spatially-varying lighting conditions of a scene can be robustly estimated from either a single, or multiple images of that scene. Second, we will show how images of objects and even full scenes can be interactively relit according to a target lighting.