Normalizing Flow-based Shape Priors for Image Analysis

Supervisor: Professor Greg Slabaugh

Shape is a fundamental property of an object.  For an object class the diversity of shapes can often be encoded into a statistical model.  This model can then be applied when analysing images that may contain instances of the object.  This project seeks to build such shape priors using normalizing flow, a generative model based on deep learning that has advantages to popular generative adversarial networks (GANs). 

The shape prior will then used in the solution of computer vision problems like image segmentation, for example using medical images that contain known anatomy (vertebrae). 

It is expected the shape prior will constrain the segmentation to achieve better results than an unconstrained method.