Unsupervised Semantic Segmentation with Language-image Pre-training

A segmentation task which does not utilise any human-level supervision for semantic segmentation except for a backbone which is initialised with features pre-trained with image-level labels.