Deep Learning for Augmented Reality

Abstract

Augmented reality aims to mix real-world visual content with virtual objects. Achieving realistic results involves solving challenging computer vision tasks, such as tracking real 3D objects and estimating the illumination conditions of a scene. In this short paper, we present how these two challenging tasks can be solved robustly and accurately with deep learning. In both cases, deep convolutional neural networks are trained on large amounts of data, and achieve state-of-the-art results.