Co-Training for Visual Object Recognition Based on Self-Supervised Models Using a Cross-Entropy Regularization.