Automatically Learned Shape and Appearance Models

Here are some example fits from a method for learning shape and appearance models. These examples are only 2D, but the framework also works for 3D image, and can be implemented in a distributed privacy-preserving fashion. No manual landmarks were used for computing the shape model. Shape was learned using a method similar to Zhang's PGA approach.

Zhang M, Fletcher PT. Bayesian principal geodesic analysis in diffeomorphic image registration. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention 2014 Sep 14 (pp. 121-128). Springer International Publishing.

The Karolinska Directed Emotional Faces

This example demostrates the Gaussian noise model on downsampled face images from the The Karolinska Directed Emotional Faces dataset. The data were fit using 40 components encoding both shape and appearance. These 40 parameters can be used for data mining.

Lundqvist, D., Flykt, A., & Ohman, A. (1998). The Karolinska Directed Emotional Faces - KDEF, CD ROM from Department of Clinical Neuroscience, Psychology section, Karolinska Institutet, ISBN 91-630-7164-9.

Original Images (randomly selected from the dataset).

Original data

Full Fit (both shape and appearance).

Full fit

Appearance model only.

Full fit

Shape model only.

Full fit

First 10 modes of variability.

First few modes

IXI dataset

This example demostrates the multinomial noise model on a single slice through segmented scans from the IXI dataset. The original scans were segmented into grey matter (shown in red), white matter (green) and CSF (blue). The data were fit using 40 components encoding both shape and appearance. These 40 parameters can be used for data mining.

Original Images (randomly selected from the dataset).

Original data

Full Fit (both shape and appearance).

Full fit

Appearance model only.

Full fit

Shape model only.

Full fit

First 10 modes of variability.

First few modes