Evaluation of Semi-supervised Semantic Segmentation for Remote Sensing, Medical Imaging, and Machine Vision Settings

Semi-supervised semantic segmentation (S4) has garnered significant attention in recent years due to the time-consuming and costly process of creating pixel-level annotations. Instead of only relying on labeled data, semi-supervised approaches leverage both labeled and unlabeled data to mitigate the issue of the labor-intense annotation process. Although current state-of-the-art methods in S4 achieve impressive results, they are often only evaluated in specific domains, which are not fully representative of many real-world applications. For this reason, we evaluate the foundational Mean Teacher approach together with UniMatch, one of the current state-of-the-art methods, on multiple datasets spanning remote sensing, medical imaging, and machine vision settings. Our results demonstrate that semi-supervised approaches are able to achieve significant performance gains in label-scarce environments and even surpass the fully supervised baseline with 100% of the labels in the machine vision setting.