For my guided research project, we present an image-based relighting method that can synthesize scenes under novel lighting using image synthesis models based on deep learning. Our method extends the Deffered Neural Rendering pipeline to perform relighting tasks. This pipeline combines the traditional graphics pipeline with learnable components: neural textures and deferred neural ren- derer. We evaluate extensively the effectiveness of our approach in several experiments, on both synthetic and real scenes. Results prove that Neural Rendering pipeline is able to reproduce complex relighting tasks like modeling high frequency lighting effects such as specularities and shadows.

Results on Light Stage Dataset

Advisor: Justus Thies Supervisor: Matthias Niessner

Github repository Document