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

Recent studies have witnessed that self-supervised methods based on view synthesis obtain clear progress multi-view stereo (MVS). However, existing rely the assumption corresponding points among different views share same color, which may not always be true in practice. This lead to unreliable signal and harm final reconstruction performance. To address issue, we propose a framework integrated with more reliable supervision guided by semantic co-segmentation data-augmentation. Specially, excavate mutual from images guide consistency. And devise effective data-augmentation mechanism ensures transformation robustness treating prediction of regular samples as pseudo ground truth regularize augmented samples. Experimental results DTU dataset show our proposed achieve state-of-the-art performance unsupervised methods, even compete par supervised methods. Furthermore, extensive experiments Tanks&Temples demonstrate generalization ability method.