Developments and applications in deep generative modelling


Charlie Nash

This research aims to develop methods for scene understanding tasks that leverage the growing supply of 3D models that are available in online databases. More specifically, the goal is to estimate scene variables such as the pose or shape of an object in a given image by training a recognition model on synthetic training data. This process would be informed by a part-based shape model which learns a probability distribution over shape variation by pre-training on a collection of 3D models.


Supervisors: Chris Williams & Vittorio Ferrari