Instance segmentation is an active topic in computer vision that is usuallysolved by using supervised learning approaches over very large datasetscomposed of object level masks. Obtaining such a dataset for any new domain canbe very expensive and time-consuming. In addition, models trained on certainannotated categories do not generalize well to unseen objects. The goal of thispaper is to propose a method that can perform unsupervised discovery oflong-tail categories in instance segmentation, through learning instanceembeddings of masked regions. Leveraging rich relationship and hierarchicalstructure between objects in the images, we propose self-supervised losses forlearning mask embeddings. Trained on COCO dataset without additionalannotations of the long-tail objects, our model is able to discover novel andmore fine-grained objects than the common categories in COCO. We show that themodel achieves competitive quantitative results on LVIS as compared to thesupervised and partially supervised methods.