The PhD qualification exam of the graduate school in Saarbrücken consists of a 25 minutes talk plus additional questions and discussion. In my case, the examiners where Prof. Bernt Schiele and Dr. Mario Fritz from the Max Planck Institute for Informatics, and Prof. Holger Hermanns from Saarland University.
In this work, we tackle the problem of 3D shape completion from partial and sparse point clouds, a fundamental problem in computer vision and computer graphics. Recent approaches can be characterized as data-driven or learning-based. The former approaches rely on pre-trained shape models whose parameters are optimized to fit the observations; this can directly be done on real data. Learning-based approaches avoid this optimization step by directly learning shape completion on synthetic data in an end-to-end fashion. However, full supervision is usually not available for real data. Therefore, we propose a weakly-supervised learning-based approach to 3D shape completion that neither requires slow optimization nor full supervision. Based on a learned shape prior, we amortize, i.e. learn, maximum likelihood fitting using deep neural networks resulting in efficient shape completion without sacrificing accuracy. We demonstrate the applicability of our approach on real and synthetic data of various object categories.