This dataset contains the parameters of a random orbit model, where the template represents the mean shape of a population and deviations from this mean shape are encoded by geodesics sampled from a multivariate Gaussian distribution.
shape_pca_template_{01234}.nii contain the template with increasingly refined levels of details.
shape_pca_subspace_scaled.nii contains a low-dimensional orthogonal subspace from which geodesic-encoding velocities are sampled.
model_variables.mat contains variables
A (100x100 matrix) - contains the covariance of the learned prior over latent variables; i.e., their distribution in the population.
Az (100x100 matrix) - contains the posterior covariance over latent variables; i.e., the uncertainty about the true value of any latent code.
lam - contains the precision (i.e., inverse of variance) of the residual noise not captured by the 100-dimensional subspace.
References
Ashburner, J., Brudfors, M., Bronik, K. and Balbastre, Y., 2019. An algorithm for learning shape and appearance models without annotations. Medical image analysis, 55, pp.197-215.
Balbastre, Y., Brudfors, M., Bronik, K. and Ashburner, J., 2018. Diffeomorphic brain shape modelling using Gauss-Newton optimisation. MICCAI 2018, pp. 862-870.