Maximizing Spatio-Temporal Entropy of Deep 3D CNNs for Efficient Video Recognition