Splitting Methods for 3D Reconstruction and SLAM

• Applicant: Professor Dr. Daniel Cremers, since 8/2020
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: from 2018 to 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 410252185

Among the central challenges in computer vision is the 3D reconstruction of the world from moving cameras. This challenge comes in many facets from classical multiple-view reconstruction, over photometric stereo, bundle adjustment and simultaneous localization and mapping (SLAM). While we have recently observed a surge of methods with increasing levels of real-world performance, we believe that there is a fundamental limitation in the optimization techniques that are currently employed in this domain: While the optimization problems are typically non-smooth and non-convex, most researchers make use of classical gradient-based techniques like gradient descent or Gauss-Newton optimization which are known to converge slowly and not well scalable towards high dimensionality. The aim of this project is to develop splitting methods to tackle the typically huge non-smooth and non-convex optimization problems in a manner that efficiently provides high-quality solutions with and scalable runtime. We will propose suitable algorithms, we will analyze their convergence properties, and we will deploy them in a multitude of applications in the domain of image-based 3D reconstruction. We expect these splitting methods to have a substantial impact on the quality and runtime of respective computer vision methods, thereby further boosting the real-world capability of respective computer vision methods.

• DFG Programme: Research Grants
• Ehemaliger Antragsteller: Dr. Tao Wu, until 8/2020