DINO
Segmented Images, No Labeled Data: Improved unsupervised learning for semantic segmentation
Training a model to separate the objects in a picture typically requires labeled images for best results. Recent work upped the ante for training without labels.