An Ensemble Neural Network for Scene Relighting with Light Classification

Liping Dong, Yu Zhu, Zhuolong Jiang, Xiangyu He, Zhaohui Meng, Chenghua Li, Cong Leng, Jian Cheng

Abstract: Illumination is a very important environmental condition. Objects in different illumination environments will present different light and shadow effects. Different kinds of illumination sources will cause different brightness and colors on the surface of the object. The conversion of illumination in two pictures is an interesting and challenging new task, which will be useful in the fields of photography and computer graphics. To solve this problem, we propose a novel solution with three stages: illumination classification, One-to-One Relighting, and Any-to-Any Relighting. Our solution can accurately classify the illumination condition of the input image and can change the direction of the illumination source from any direction to another. We evaluate our methods on VIDIT, a rendered dataset of artificial scenes. The proposed solution produces good results under different light conditions.