Three-Dimensional Human Body Reconstruction Based on Volumetric Rendering Learning

Currently, three-dimensional human body reconstruction technology plays a significant role in various domains related to everyday life, such as virtual reality, augmented reality, and e-commerce. In previous studies, Neural Radiance Fields (NeRF) have been commonly used to represent three-dimensional human body models. However, due to the lack of surface constraints on the neural radiance field, the obtained models often exhibit rough surfaces and suffer from adhesion. Therefore, in order to obtain three-dimensional human body models with fine surface details, we propose a method for three-dimensional human body reconstruction based on volumetric rendering learning. We represent the surface of the human body model as the zero level set of a Signed Distance Function (SDF) and introduce a novel volumetric rendering method called Occlusion-Aware Unbiased Weighting Function to train the SDF representation. The signed distance function (SDF) naturally regularizes the learned geometric shape of the human body, while the occlusion-aware unbiased weighting function effectively reduces inherent geometric errors during the reconstruction process, enabling high-quality reconstruction of three-dimensional human bodies. We validated the effectiveness of our proposed method on the PeopleSnapshot dataset. Meanwhile, considering the high prevalence of adolescent scoliosis and the practical need for scoliosis brace modeling, we integrated the reconstruction of the three-dimensional human body with the production of a scoliosis brace model to improve the fit between the scoliosis brace model and the patient.