Improving photometric quality of SLAM

Abstract

We offer a Master Thesis in the topic of 3D computer vision and deep learning, in particular on real-time 3D reconstruction from a sequence of frames. Existing SLAM approaches rely on different fusion methods to integrate sensor data from different view angles in order to reconstruct a scene. For example, KinectFusion^?[1] uses running average on TSDF[3] and RGB values on each voxel. Alternative fusion methods were proposed [2]. These methods reconstruct the geometry accurately, nevertheless the reconstructed texture is usually blurry and and missing fine details (See Figure1).

In this project, we aim to develop a novel approach that is able to reconstruct a more photorealistic scene in real-time. We are looking for a motivated student with prior experience in deep learning and computer vision, and good knowledge of Python and C++. The project will be carried out in cooperation with Google Zurich.

Requirements

• Prior experience (theoretical and practical) of computer vision and deep learning.
• Knowledge of SLAM methods.
• Good knowledge of Pytorch
• Good coding skills in python and C++

Literature

• [1] Izadi, Shahram, et al. "KinectFusion: real-time 3D reconstruction and interaction using a moving depth camera." Proceedings of the 24th annual ACM symposium on User interface software and technology. 2011.
• [2] Keller, Maik, et al. "Real-time 3d reconstruction in dynamic scenes using point-based fusion." 2013 International Conference on 3D Vision-3DV 2013. IEEE, 2013.
• [3] S. Osher and R. Fedkiw. Level set methods and dynamic implicit surfaces. Springer, 2002.

Application

If you are interested, please contact me via e-mail:

• Shun-Cheng Wu