OpenCoS: Contrastive Semi-supervised Learning for Handling Open-Set Unlabeled Data

Semi-supervised learning (SSL) has been a powerful strategy to incorporate few labels in better representations. In this paper, we focus on practical scenario that one aims apply SSL when unlabeled data may contain out-of-class samples - those cannot have one-hot encoded from closed-set of classes label data, i.e., the is an open-set. Specifically, introduce OpenCoS, simple framework for handling realistic semi-supervised based upon recent self-supervised visual representation learning. We first observe open-set dataset can be identified effectively via contrastive Then, OpenCoS utilizes information overcome failure modes existing state-of-the-art methods, by utilizing pseudo-labels and soft-labels in- respectively. Our extensive experimental results show effectiveness under presence samples, fixing up methods suitable diverse scenarios involving data.