CSci 5563 Final Project Spring 2021
Group Members: Luis Guzman, Isaac Kasahara, Aditya Rajguru, and Helena Shield
Abstract: Relighting images and videos has proven to be a very difficult task in the field of computer vision. With augmented reality requiring objects placed in scenes to match the scene's lighting, and movies requiring perfect lighting for each shot, the ability to change a video's lighting after it has been recorded is quickly becoming a relevant task. Current relighting methods have a heavy focus on faces, and fail to consider the challenges that video relighting provides. In this paper we demonstrate a method that takes advantage of scene geometry by estimating the normals of an image, uses existing networks to generate the reflectance image, and utilize Lambertian shading to generate the image under a new lighting condition. We utilize averaging between frames to reduce flicker and enforce continuity. Although our method produced higher error values than the baseline, we believe that our method produces qualitatively better results when evaluating geometrically consistent highlights and the reduction of flicker in video.
![](https://i.ytimg.com/vi/tt2wKUsNhy4/maxresdefault.jpg)