Illumination Estimation Using ML

Abstract

This research proposes another way of determining light source position using Machine Learning. In the century that promises rise and rapid growth of Artificial Intelligence (AI) we can expect that computational capabilities of our machines will enable us to embed such AI-driven solutions in almost any software and/or hardware solution. Another promising trend in technologies is Augmented Reality (AR). Today we already can see that our mobile devices with AR on board can deliver significant amount of user-driven content. The good examples that generate big revenues and have millions of daily users are AR filters that allow users to place a virtual object on top of the face. Big players in this niche are Snapchat and Instagram, I think that everyone smiled at least once when using one of their filters. Following the trends, I took the liberty of exploring the possibility of illuminating virtual object depending on user’s face illumination. The idea is to let robust and compact Neural Network decide about how to cast light on virtual component while being efficient enough to be deployed on mobile device. Having limited resources this thesis explores only the tip of the great potential of such technology.