Pose-Driven Compression for Dynamic 3D Human via Human Prior Models.

Area of Science:

• Computer Vision
• Computer Graphics
• Multimedia Systems

Background:

• Efficient compression of dynamic 3D human data is essential for immersive multimedia applications.
• Existing methods often focus on signal-level errors, limiting perceptual quality and downstream task support.
• A novel approach leveraging human priors is needed for advanced 3D human representation.

Purpose of the Study:

• To develop a cost-effective and high-quality dynamic 3D human compression framework.
• To enhance perceptual quality and support diverse downstream applications like visual analysis and content editing.
• To outperform existing 3D codecs in subjective quality and bitrate efficiency.

Main Methods:

• A layered coding architecture is proposed, starting with a pose-driven Avatar as the structure layer.
• Human motion is parameterized using the Skinned Multi-Person Linear Model (SMPL) for the motion layer.
• Normals are incorporated as an enhancement layer to preserve fine geometric details.
• The Avatar, SMPL parameters, and normal maps are compressed into semantic bitstreams.

Main Results:

• The proposed framework demonstrates remarkable outperformance against state-of-the-art 3D codecs.
• Significant improvements in subjective quality are achieved with minimal bitrate consumption.
• The framework's superiority in perceptual quality and bitrate savings increases with sequence size and frame count.

Conclusions:

• The human prior-based compression framework offers a significant advancement in dynamic 3D human data transmission.
• The layered approach effectively balances compression efficiency with high perceptual quality and functional versatility.
• This method provides a promising solution for real-time, high-fidelity 3D human rendering in various applications.