Beyond Features: Dense and Direct Methods for Visual SLAM and Geometric Reconstruction

Daniel Cremers (TU Munich, Germany)

Abstract:

The reconstruction of the 3D world from a moving camera is among the central challenges in computer vision. While traditional approaches have been focused on computing correspondence and 3D structure for a sparse set of feature points, more recent approaches aim at directly computing dense geometric surfaces using all available image data. In my talk, I will present some recent developments on convex formulations for dense reconstruction from multiple images or multiple videos. Furthermore, I will present real-time capable direct methods for reconstructing the world from handheld color or RGB-D cameras. Applications include free-viewpoint television and augmented reality.

This is joint work with Kalin Kolev, Martin Oswald, Jakob Engel, Jan Stühmer, Christian Kerl, Frank Steinbrücker and Jürgen Sturm