Uncertainty-Aware Stratified Scribble Propagation for Weakly Supervised Human Skin Segmentation

Segmenting human skin regions in digital images is a well-studied computer vision problem with a wide range of applications. State-of-the-art deep learning models achieve high accuracy, but their training requires large sets of precisely annotated pixel-level masks, whose preparation remains a time-consuming and tedious task. In this paper, we propose a weakly supervised framework that trains competitive skin segmentation networks from sparse, single-class scribble annotations. We exploit the concept of a well-established skin-aware shortest-path propagation scheme, adapting it to generate tri-level pseudo-labels consisting of high-confidence skin and background regions, and an uncertain region that is excluded from the training loss. We evaluate the framework on the ECU benchmark using two architectures: the lightweight Skinny convolutional neural network and the transformer-based SegFormer-B2. The proposed method achieves an F-score of 0.85 on Skinny using the reproducible simulated-scribble protocol and 0.87 using manually drawn scribbles. When the model is additionally fine-tuned on just 80 precisely annotated images, F-score improves to 0.89, approaching fully supervised accuracy while requiring substantially fewer dense annotations. The proposed pseudo-labeling framework also outperforms a content-agnostic ScribbleSup baseline, and the ECU-trained models transfer to the HGR hand-gesture dataset without any adaptation. While we evaluate the framework on skin segmentation, its underlying principles may be applicable beyond this specific task and point to broader potential in weakly supervised segmentation.