Generative Inpainting for Hand Detection in Occupational Safety

A recent paper on arXiv (2606.01896) explores how generative inpainting can enhance hand detection models for safety in the workplace. This research tackles the issue of distribution shifts in public datasets, which predominantly feature bare hands and fail to adequately represent variations such as gloves, tattoos, and jewelry. By modifying only the hand area in actual images to add accessories, the authors assess if synthetic data can bridge this gap. They conduct six experiments (A-F) to evaluate YOLOv8n hand detectors, with four focused on training (A, C, D, E) using three random seeds each, utilizing both real and synthetic paired datasets. The full paper can be accessed at https://arxiv.org/abs/2606.01896.