Disentangled Human Body Embedding Based on Deep Hierarchical Neural Network.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Computer Graphics

Background:

• Human body shapes vary significantly with identity and pose.
• Existing methods struggle to effectively disentangle shape and pose in 3D.
• Latent representation learning offers potential for compact body modeling.

Purpose of the Study:

• To develop a disentangled latent space for 3D human body shape and pose.
• To improve the accuracy and flexibility of 3D human body reconstruction and generation.
• To create a robust embedding applicable to diverse body shapes and poses.

Main Methods:

• An autoencoder-like network architecture was designed for disentangled embedding.
• A hierarchical reconstruction pipeline was implemented to enhance disentangling.
• A large dataset of human body models with consistent connectivity was constructed.
• Deformation-based latent representation learning principles were applied.

Main Results:

• The learned embedding achieved superior reconstruction accuracy for 3D human bodies.
• The model demonstrated high flexibility in 3D human body generation through interpolation and sampling.
• Extensive experiments validated the effectiveness of the proposed embedding.

Conclusions:

• The proposed method effectively disentangles 3D human body shape and pose.
• The learned embedding offers significant advancements in 3D human body modeling and generation.
• This approach has broad applicability in areas like animation, virtual reality, and medical imaging.