HYVE: Hybrid Vertex Encoder for Neural Distance Fields.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Neural Networks

Background:

• Neural shape representation uses neural networks to encode 3D geometry.
• Existing methods often require latent-code optimization, which is computationally intensive.

Purpose of the Study:

• To present a neural network architecture for accurate 3D shape encoding in a single forward pass.
• To improve surface reconstruction quality and computational efficiency.

Main Methods:

• A multi-scale hybrid system combining graph-based and voxel-based components.
• Novel projection of point-based features onto a grid, insensitive to point density.
• Training network to solve the eikonal equation, using zero-level set for training and inference.
• Modification of loss function for non-well-defined surface normals.

Main Results:

• Achieved smoother and more detailed 3D shape reconstructions, especially with oriented point clouds.
• Outperformed existing baselines in surface reconstruction across various datasets.
• Demonstrated higher computational efficiency and reduced parameter count.

Conclusions:

• The proposed neural network architecture offers an efficient and effective approach to 3D shape representation and reconstruction.
• The hybrid system and eikonal equation-based training provide robust and accurate results, even for complex geometries.