Curiosity-driven 3D Scene Structure from Single-image Self-supervision
David Griffiths, Jan Boehm, Tobias Ritschel
arXiv preprint, 2020
We present a novel method for fully self-supervised scene paramaterisation from a single image. We achieve this by employing analysis-by-synthesis using a differentiable renderer. We show that a simple L2 loss is not sufficient for such a task. Instead, we introduce a GAN-like critic to constrain the network to propose realistic outputs. By adding such a constraint we observe the L2 loss is now sufficient to solve a number of tasks on both synthetic and real data.