VFIR: Vector Fields Implicit Representation for Surface Reconstruction From Point Clouds.

Area of Science:

• Computer Vision
• 3D Shape Modeling
• Geometric Deep Learning

Background:

• Surface reconstruction from point clouds is a key computer vision challenge.
• Existing deep learning methods using signed/unsigned distance functions have limitations (watertightness, non-differentiability).

Purpose of the Study:

• Introduce VFIR, a novel neural implicit function approach for 3D shape modeling using vector fields (VFs).
• Improve fitting accuracy and learning smoothness in surface reconstruction.

Main Methods:

• Leveraging neural implicit functions to learn vector fields (VFs).
• Enhancing fitting by displacing points along predicted vector directions.
• Employing a progressive learning strategy with denser point clouds and normals.
• Utilizing truncated vector fields (TVFs) and an optimized three-state marching cubes (OT-MC) algorithm for isosurface extraction.

Main Results:

• VFIR demonstrates state-of-the-art performance in surface reconstruction.
• Achieves high accuracy and robustness across diverse 3D models.
• Exhibits strong generalization capabilities.

Conclusions:

• VFIR offers a promising solution for 3D shape modeling and surface reconstruction.
• Overcomes limitations of traditional distance function-based methods.
• Presents a robust and accurate approach applicable to various real-world scenarios.