Pixel-wise semantic segmentation on different domains is a well-researched area that shows satisfactory results when supervision is provided. However, supervision for such methods is only available for a small subset of real-world scenarios. Lack of supervision for real-world scenarios requires new methods that can work directly on input images without supervision. Unsupervised segmentation methods provide a solution for such cases.

We experimented with a domain where supervision is not available. Our domain is the black-and-white microscopy images of Arbuscular mycorrhizal, a symbiotic relationship between soil fungi and vascular land plants. We segment the nutrition-transferring instances in the symbiotic structure inside the plants. We perform this segmentation with clustering on different embedding spaces that are extracted from the input images only. We use Visual Transformer-based self-supervised feature extractors to obtain a meaningful representation and cluster the pixels on the obtained embedding space. We show that our method outperforms the baseline method of color-based clustering. Finally, we trained a UNet model to experiment with different levels of supervision and compare how the fully unsupervised method performs compared to different levels of supervision.