3D Computer Vision in the Age of Deep Learning

Daniel Cremers

Technical University of Munich, DE

Abstract

In my presentation, I will discuss both optimization-based and learning-based approaches to 3D reconstruction and simultaneous localization and mapping (SLAM). I will start with traditional optimization-based approaches to 3D reconstruction and SLAM, discuss differences between keypoint-based and direct approaches, including classical methods like ORB SLAM, LSD SLAM and DSO. Then I will discuss how learning-based formulations can help us boost the performance of traditional methods, providing more precision, more robustness and ultimately dense reconstructions from a moving camera with methods like MonoRec or TANDEM. While most of my talk is focused on reconstructions of the static world, toward the end I will show recent developments on reconstructions of dynamic scenes with methods like AnyCam or 4Deform. If time permits, I will also discuss neural methods for 3D shape analysis.