Parametric Body Reconstruction Based on a Single Front Scan Point Cloud.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Digital Human Modeling

Background:

• Current 3D body scanning systems are often bulky, expensive, and require specific clothing, limiting their accessibility and application.
• Acquiring accurate digital body models from clothed individuals presents significant challenges due to occlusions and complex garment interactions.

Purpose of the Study:

• To develop an efficient pipeline for reconstructing accurate digital body models from a single front-view scan of a clothed person.
• To overcome the limitations of existing 3D scanning technologies by enabling reconstruction without requiring specialized clothing or multiple scans.

Main Methods:

• A novel pipeline combining inner body shape inference and parametric model registration is proposed.
• Three neural network modules (Scan2Front-Net, Front2Back-Net, Inner2Corr-Net) predict front inner, back inner, and reference point clouds.
• The method learns axial offset relationships within 14 body parts to infer the back inner shape from the front scan.

Main Results:

• The proposed pipeline successfully reconstructs detailed digital body models from single front scans of clothed individuals.
• Qualitative and quantitative analyses demonstrate significant improvements in body shape completion and reconstruction accuracy compared to existing methods.
• The approach effectively infers the inner body shape, enabling accurate parametric model registration.

Conclusions:

• The developed method offers a practical and accurate solution for digital body model reconstruction from single-view scans of clothed subjects.
• This approach has the potential to advance applications in areas such as virtual try-on, personalized fashion, and biomechanics.
• The pipeline demonstrates robustness and accuracy, addressing key challenges in clothed human body modeling.