Machine Learning for Three-Dimensional Reconstruction and Visual Localization

The lecture will be followed by a discussion


ABSTRACT
Reconstructing objects and scenes from a set of images (the 3D reconstruction
problem), as well as estimating the position and orientation from which images
were taken (the visual localization problem), are fundamental problems in
computer vision, with applications in autonomous driving, augmented reality and
robotics, among others. In this talk, we will discuss recent advances in both 3D
reconstruction and visual localization that have been made possible through
machine learning. Starting with the 3D reconstruction problem, this talk will
first discuss recent state-of-the-art algorithms to produce (close-to)
photorealistic scene representation by estimating so-called neural
radiance fields. The second part of the talk will focus on the visual
localization problem. In particular, we will discuss using accurate (and
less-accurate) 3D models to allow robust camera position and orientation
estimation. In addition, we will discuss privacy aspects of the visual
localization problem.