Self-supervised Multi-view Stereo via Effective Co-Segmentation and Data-Augmentation