Fig.2:

Data augmentation: PCA is used to compute the PCA scores ${\left\{z_n\right\}}_{n=1}^N$ of the training shape samples ${\left\{x_n\right\}}_{n=1}^N$. Augmented samples z_s are randomly drawn from a collection of multivariate Gaussian distributions $K_{\sigma }^n(z)$, each with a training example z_n as the mean and covariance σ²I_L. The correspondences x_s are used to compute a TPS warp to map x_n to x_s, which is used to warp the respective image y_n to a new image y_s with known shape parameters z_s.