Efficiently Learning a Detection Cascade with Sparse Eigenvectors