Fachbereich Informatik - Aktuell

06.09.2017 11:24

Disputation Benjamin Resch

am Dienstag, 19. September 2017 in Raum A301, Sand 1, 2. OG.

Robust and Efficient Camera-based Scene Reconstruction

Berichterstatter 1: Prof. Dr.-Ing. Hendrik P. A. Lensch
Berichterstatter 2: Prof. Dr. Andreas Schilling

For the simultaneous reconstruction of 3D scene geometry and camera poses from images or videos, there are two major approaches: On the one hand it is possible to perform a sparse reconstruction by extracting recognizable features from multiple images which correspond to the same 3D points in the scene. With those features, the positions of the 3D points as well as the camera poses can be obtained such that they explain the positions of the features in the images best. On the other hand, on video data, a dense reconstruction can be obtained by alternating between the tracking of the camera pose and updating a depth map representing the scene per frame of the video.
In this dissertation, we introduce several improvements to both reconstruction strategies. We start from improving the reliability of image feature matches which leads to faster and more robust subsequent processing. Then, we present a sparse reconstruction pipeline completely optimized for high resolution and high frame rate video, exploiting the redundancy in the data to gain more efficiency. For (semi-)dense reconstruction on camera rigs which is prone to calibration inaccuracies, we show how to model and recover the rig calibration online in the reconstruction process. Finally, we explore the applicability of machine learning based on neural networks to the relative camera pose problem, focusing mainly on generating optimal training data.
Robust and fast 3D reconstruction of the environment is demanded in several currently emerging applications ranging from set scanning for movies and computer games over inside-out tracking based augmented reality devices to autonomous robots and drones as well as self-driving cars.