Deep Consistent Illumination in Augmented Reality

Xiang Wang, Kai Wang, Shiguo Lian

Abstract: Visual coherence is a basic requirement for providing a realistic scene in Augmented Reality. As one of the most important factors, illumination consistency is usually achieved through two steps: real lighting estimation and virtual object rendering. In this paper, we propose an end-to-end method to adjust the luminance and shadow of the virtual object in the composite image using a deep neural network. It learns the varying appearances of the virtual object under different illumination and viewpoints from a large amount of synthesized images, and transforms the illumination of the composite image into coherence according to the appearance of the real world. The effectiveness of our method is demonstrated through extensive experiments.

External IDs:dblp:conf/ismar/WangWL19