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

Area of Science:

• Biomedical Engineering
• Computer Vision
• Anthropometry

Background:

• 3D body scans offer accurate body shape measurement but struggle with detailed body fat analysis due to mesh complexity.
• Current methods often rely on indirect anthropometric measurements, missing intricate body shape details.
• Existing point-based methods lack focus on human body shape and regression tasks.

Purpose of the Study:

• To explore the feasibility of using point cloud representations for body fat percentage analysis from 3D body scans.
• To introduce and evaluate a novel transformer-based model (D3BT) for accurate body fat prediction.
• To improve upon traditional anthropometric and existing point-based approaches in body composition assessment.

Main Methods:

• Developed D3BT, a transformer network operating on 3D body point clouds.
• Implemented dynamic voxelization for enhanced point cloud density and alignment.
• Trained and evaluated models using ground truth data from dual-energy X-ray absorptiometry (DXA).

Main Results:

• The D3BT model demonstrated superior performance in predicting body fat percentage compared to traditional and other point-based methods.
• Achieved an average R-squared score of 0.86, indicating strong predictive accuracy.
• Reduced Root Mean Square Error (RMSE) by an average of 10.30%.

Conclusions:

• The D3BT model effectively leverages point cloud data from 3D body scans for precise body fat estimation.
• Transformer networks offer a promising avenue for analyzing complex human body shape data.
• This approach advances body composition assessment beyond traditional anthropometric measurements.