A generative model, as the name suggests, generates data based on a parameter. 3D generative models parametrically generate 3D data from a learned distribution of 3D shapes. The distribution is often learned from a 3D dataset. This area of research is now highly active due to the explosion of research in neural generative models such as variational auto-encoders, and generative adversarial networks.

One of the prominent use cases of 3D generative models is reconstruction from images and videos, where these models are used as priors. They are also useful for scene understanding and working with raw 3D data.

In this seminar, we discuss various 3D generative models starting from classical PCA-based models to state-of-the-art neural network-based models. We discuss generative models based on different 3D shape representations - mesh, and implicit representations.

• Organizers: Tarun Yenamandra, Marvin Eisenberger, Maolin Gao
• Email: 3dgm-ws21@vision.in.tum.de (please address all requests to this address only!)
• Due to the current situation with the Coronavirus, the preliminary meeting will be online. The seminar and meetings with the supervisor should be in person. We hope that the situation will not get worse and that we do not have to switch back to the online format.

Pre-Meeting Slides: slides

Zoom Link for the Seminar (password protected)

• Day: Tuesday
• Time: 10:00 - 12:00
• Location: Zoom
• Each session will consist of two talks
• All talks are held in English