Learning Canonical $\mathcal{F}$-Correlation Projection for Compact Multiview Representation