Computational illumination for portrait photography and inverse graphics

Abstract

Supervised training of deep networks has led to remarkable successes in computer vision, for example on image classification or object detection problems. These successes are driven by the availability of large amounts of paired training data with manual ground truth annotations. For many photography or inverse graphics applications however, manual annotation of ground truth labels is not viable. Motivated by this, the research presented in this thesis proposes several portable hardware prototypes that enable the collection of training data for applications ranging from non-line-of-sight imaging to relighting and dark-flash photography.

The thesis also discusses a novel formulation for fast and accurate differentiable rendering based on analytical anti-aliasing. It is demonstrated how this renderer can be used for inverse graphics problems. The thesis concludes with a discussion on how differentiable programming can be combinded with data-driven feed forward networks for practicle inverse graphics applications.