These sequences of landmarks will now form the shape outlines, and a given training shape can be defined as a vector. We generally assume this scattering is Gaussian in this space, and we use PCA to compute normalized eigenvectors and eigenvalues of the covariance matrix across all training shapes. Using the top-center eigenvectors, we create a matrix of dimensions 2k * m, which we will call P. This way, each eigenvector describes a principal mode of variation along the set.