Visual Localization

Torsten Sattler

Chalmers University of Technology, SE

Abstract

Visual localization is the problem of estimating the exact position and orientation from which an image was taken with respect to a known scene.
Visual localization plays a central part in many exciting applications of computer vision, including Augmented / Mixed / Virtual Reality and autonomous robots such as self-driving cars.

This lecture covers the topic of visual localization in two parts: The first part provides an overview over the topic by first discussing classical feature-based localization pipelines and then introducing different ways of integrating machine learning into such pipelines.

The second part first discusses the short-comings of existing localization systems, especially in the context of handling complex scenes whose appearance changes over time. I will then discuss current efforts on integrating higher-level scene understanding into the localization process to make localization algorithms more robust and reliable.