sSfS: Segmented Shape from Silhouette Reconstruction of the Human Body

Area of Science:

• Computer Vision
• Medical Imaging
• Biometrics

Background:

• Accurate 3D human body shape estimation is crucial for diverse applications like medicine and special effects.
• Existing methods, such as Shape from Silhouette (SfS), face challenges in reconstructing complex human poses and concave regions.
• The demand for high-quality, complete, and accurate human body models is steadily increasing.

Purpose of the Study:

• To present a novel approach, segmented Shape from Silhouette (sSfS), that enhances conventional voxel-based SfS for 3D human body reconstruction.
• To improve the accuracy and quality of 3D human body shape estimation, especially in challenging concave areas.
• To validate the effectiveness of the sSfS method against state-of-the-art SfS using a comprehensive dataset.

Main Methods:

• The study extends the conventional voxel-based Shape from Silhouette (SfS) method by incorporating silhouette segmentation.
• The proposed segmented Shape from Silhouette (sSfS) method enables separate 3D reconstruction of individual body segments.
• A validation dataset featuring the human body in 20 complex poses was utilized, with assessments based on quality metrics against photogrammetric ground-truth.

Main Results:

• The sSfS method demonstrated superior performance in 3D human body shape estimation compared to the standard SfS approach.
• The number of invalid reconstruction voxels was 1.7 times lower using sSfS than with the state-of-the-art SfS.
• The root-mean-square (RMS) error for distance to the reference surface was reduced by a factor of 1.22 with sSfS.

Conclusions:

• The novel sSfS method offers significantly improved 3D human body shape estimation, particularly excelling in reconstructing concave body areas.
• sSfS provides more accurate and complete human body models, addressing limitations of previous SfS techniques.
• The findings suggest sSfS is a promising advancement for high-quality 3D human body modeling in various fields.