Abstract

In graphics, we have sophisticated algorithms that simulate light interactions to generate photorealistic images from scene descriptions. How should we invert these algorithms to recover parameters from the induced images (i.e., infer cause from effect)? I will present our recent efforts in differentiating these algorithms and using the resulting gradients to solve inverse problems in the spirit of analysis-by-synthesis. Focusing specifically on recovering the geometry of the scene, I will discuss our work on addressing the challenges of differentiating and designing rendering objectives with respect to scene parameters.