MAGIC: Learning from Visibility Asymmetry for Unsupervised Stereo Matching
S3 is a novel unsupervised stereo matching framework that uses multi-baseline geometry consistency and viewpoint asymmetry to create reliable pseudo-labels, even when parts of the target view are occluded. It assigns different target images to teacher and student networks, allowing occluded regions in the student's view to be visible to the teacher for better supervision.