D3BT: Dynamic 3D Body Transformer for Body Fat Percentage Assessment

3D body scan has been adopted for body composition assessment due to its ability to accurately capture body shape measurements. However, the complexity of mesh representation and the lack of fine-shape descriptors limit its applications in body fat percentage analysis. Most studies rely on algorithms applied to anthropometric values derived from 3D scans, such as multiple girth measurements, which fail to account for the body's detailed shape. To address these issues, we explore the feasibility of using point cloud representation. However, few existing point-based methods are aimed at the human body or regression tasks. In this study, we introduce a new model, D3BT, which utilizes a transformer-based network on the body point cloud to efficiently learn shape information for regional and global fat percentage regression tasks. The model dynamically divides the points into voxels for enhanced transformer training, providing higher density and better alignment across different subjects, which is more suitable for body shape learning. We evaluate various models for predicting body fat percentage from 3D body scans, using ground truth data from dual-energy X-ray absorptiometry (DXA) reports. Compared to traditional methods that depend on anthropometric measurements and other point-based approaches, the proposed model shows superior results. In extensive experiments, the model reduces the Root Mean Square Error (RMSE) by an average of 10.30% and achieves an average R-squared score of 0.86.