TACoS: Weakly Supervised Learning of Two-Dimensional Materials from Scribble Annotations to Precise Segmentation
TACoS is a new method for precisely segmenting 2D materials using very few annotations (scribbles). It combines consistency learning, tree energy constraints, and contrastive learning to learn detailed boundaries even with sparse labels. The result achieves over 96% of full supervision performance with less than 0.6% annotated data.