This work shows how self supervised features and over clustering can solve a hard imbalanced binary task with simple downstream logic. It turns limited labels into high quality pixel level decisions that nearly match a supervised approach. The paper is available online at Science Direct
