Lightweight Explicit 3D Human Digitization via Normal Integration.

Area of Science:

• Computer Vision
• 3D Computer Graphics
• Machine Learning

Background:

• 3D human reconstruction from images is crucial for various applications.
• Large neural networks pose challenges for deployment on resource-constrained edge devices due to high computational demands.
• Optimizing network architecture is key to reducing computational cost and improving efficiency.

Purpose of the Study:

• To propose a lightweight and efficient 3D human reconstruction model.
• To balance reconstruction accuracy with computational cost for practical applications.
• To enable deployment of 3D human modeling on edge devices.

Main Methods:

• Integration of Dilated Convolutions and Cross-Covariance Attention mechanism for a lightweight generative network.
• Utilizing a novel loss function designed for normal map geometric properties to enhance surface reconstruction quality.
• Developing a network architecture that captures multi-scale information while minimizing model complexity.

Main Results:

• Achieved approximately 80% reduction in training parameters compared to existing methods.
• Maintained high quality of generated 3D human models.
• Demonstrated improved surface reconstruction accuracy through the specialized loss function.

Conclusions:

• The proposed lightweight model offers an efficient solution for 3D human reconstruction.
• The integration of specific architectural components and a tailored loss function significantly reduces computational requirements.
• This advancement facilitates the practical deployment of 3D human modeling technologies on edge devices.