SCPNet: Unsupervised Cross-modal Homography Estimation via Intra-modal Self-supervised Learning.