Reality models for efficient registration in Augmented Reality

The fundamental challenge of vision-based Augmented Reality is to make virtual objects and the real world coincide in a seamless way. In order to solve this registration problem, at least one real object present in the scene has to be detected, recognized and used by a corresponding algorithm to compute the position and orientation of the camera.

In this thesis, we take a closer look at the representations of reality that can be used in the context of Augmented Reality. We define a Reality Model as a set of compact information about an object of interest and an algorithm that uses this information for estimating the pose of the camera.

We first review existing Augmented Reality systems from the standpoint of these models and propose a classification based on the requirements of the pose estimation. We then present our novel approaches for different types of reality models, with applications in marker-based and texture-based pose estimation, 3D reconstruction, object detection and learning-based methods for local pose estimation.