This paper proposes SemiOVS, a semi-supervised learning-based semantic segmentation framework that leverages limited labeled data and abundant out-of-distribution (OOD) unlabeled data. While previous studies have shown promising results using limited segmentation of standard datasets, the potential of leveraging large-scale unlabeled images has not been explored. SemiOVS utilizes the Open-Vocabulary Segmentation (OVS) model to generate highly accurate pseudo-labels for OOD images. Experimental results on the Pascal VOC and Context datasets demonstrate that leveraging additional unlabeled images in a label-constrained environment improves performance, particularly when leveraging OOD images via the OVS model. SemiOVS achieves state-of-the-art performance, outperforming existing methods PrevMatch and SemiVL by +3.5 mIoU and +3.0 mIoU, respectively.
