S2MGen: A synthetic skin mask generator for improving segmentation

Skin segmentation is an important and challenging task which finds use in direct applications such as image editing and downstream tasks such as face detection or hand gesture recognition. However, the availability of diverse and high-quality training data is a major challenge. Annotation of dense segmentation masks is an expensive and time consuming process. Existing skin segmentation datasets are often limited in scope: they include datasets built for specific downstream tasks that were captured under controlled conditions, with limited variability in lighting, scale, ethnicity, and age. This lack of diversity in the training data can lead to poor generalization and limited performance when applied to real-world images. To address this issue, we propose a tunable pipeline for dataset generation, the Synthetic Skin Mask Generator (S2MGen), which allows for the creation of dataset which includes a diverse range of body positions, camera angles, and lighting conditions. We explore the impact of these tunable parameters on skin segmentation performance. We also show that improvements can be made to the performance and generalizability of models trained on real world datasets, by the inclusion of synthetic data in the training pipeline.