Generating Masked Human Body Skeletons with Diffusion Model

Human pose estimation constitutes the foundation of computer vision. However, practical applications consistently encounter a pivotal challenge, which revolves around the accurate annotation of key skeletal points when they are obscured or missing. To address these challenges, this paper proposes an innovative solution based on a diffusion model. The primary objective is to enhance the precision of human pose prediction by employing a generative AI model to accurately predict occluded or missing skeletal points, using a fixed-size mask overlay applied to regions with the densest joint nodes. The concept of this model conceptualizes the process of predicting human poses as a denoising procedure, where complete joint point data is considered the ground truth for human poses, while missing skeletal points are treated as noise. Experimental results show that identifying regions with dense joint points and applying the diffusion model not only effectively mitigates this noise but also yields more precise human pose prediction results.