Enrico Meloni (University of Siena)

When: Feb 10, 2021 – 11:00 – 11:45 AM

Where: Google meet link

Description

Neural Models can be efficiently used for 2D and 3D reasoning. However, most 3D estimation models rely on supervised training, which comes with costly annotations of all observable 3D properties. Differentiable rendering enables 2D images to be used as supervisions for the 3D properties of the scene. This enables the use of available datasets and cheaper annotations. Differentiable rendering allows to seamlessly insert the rendering process in the learning pipeline and let gradients flow through it. This allows to study problems such as Object Reconstruction, Face Reconstruction, Body Pose Estimation, and 3D Adversarial Learning.