LGDV Research

This page describes the work of the group of Marc Stamminger.
Click here to learn about the work of Bernhard Egger’s group.

Point-Based Novel View Synthesis and Reconstruction

Point-Based scene representations are a powerful scene representation, both for novel view synthesis and 3D reconstruction. Their lack of geometry enable fast scene optimization, scene completion, and rendering, but at also present particular challenges. Our research focuses on real-time methods using such point-based representations for different applications in visual computing, from novel view synthesis to 3D reconstruction.

Real-Time 3D Reconstruction

Reconstructing 3D models from images is a long standing research problem. The wide availability of depth cameras and advances in deep learning have pushed the field enormously in the last years. Our group’s research in this area is focused on real-time methods, that reconstruct a 3D-scene on the fly from video and/or depth streams.

Imaging Radar – Forward and Inverse Simulation

Radar uses light with invisible frequencies. Imaging radars use arrays of sending and receiving antennas to reconstruct images of their environment. Simulating such radars has many similarities with physically based rendering, but there are also major differences. So, for radar phase is more important than amplitude or brightness, polarization plays an important role, and image formation is totally different from normal cameras. In our research, we look at efficient algorithms for the simulation of such imaging radars, and at its inversion for advanced 3D scene geometry and material reconstruction.

Virtual and Extended Reality – Mixing Real and Digital Worlds

In virtual reality the user is immersed into a virtual environment, in extended reality, real and digital worlds converge. In our research, we look at different variants of virtual and extended reality. We explore real-time rendering algorithms for virtual reality, look at projection mapping techniques to augment real-world objects with virtual content, and develop novel setups to augment live stage performances.

Visual Computing for Medical Applications

Imaging is a pillar of modern medicine, so visual computing has many applications in medicine. In our group, we explore visual computing approaches to support such applications, in particular human motion capture and analysis, as well as methods to reconstruct oscillating vocal folds.

Digital Agriculture

In the area of digital agriculture, the Fürth branch of our group developed methods for the generation of a digital twin of orchard trees using latest novel-view synthesis and neural rendering and reconstruction techniques.

LiDAR Resimulation

Realistic sensor simulation is mandatory for the design, evaluation, test, and training of perception for autonomous driving. We research the resimulation of LiDAR sensors, where we derive the characteristics of particular LiDAR from real test drives and generate a digital sensor twin that can be used in virtual driving environments like CARLA.