Efficiently training a better visual detector with sparse eigenvectors