CAP-UDF: Learning Unsigned Distance Functions Progressively From Raw Point Clouds With Consistency-Aware Field...

Area of Science:

• 3D Computer Vision
• Geometric Deep Learning
• Surface Reconstruction

Background:

• Current surface reconstruction methods often rely on signed distance functions, limiting them to closed surfaces.
• Unsigned distance functions (UDFs) can represent open surfaces but struggle with smoothness due to point cloud discontinuities.
• Existing UDF learning methods face challenges in producing smooth distance fields from raw point clouds.

Purpose of the Study:

• To propose CAP-UDF, a novel method for learning consistency-aware unsigned distance functions (UDFs) from raw point clouds.
• To enable accurate reconstruction of both open and closed surfaces.
• To improve the smoothness and accuracy of learned distance fields for point cloud data.

Main Methods:

• Learning consistency-aware UDFs by enforcing a field consistency constraint on query points.
• Training a neural network to dynamically infer query-to-surface relationships for progressive surface estimation.
• Utilizing a novel polygonization algorithm based on learned UDF gradients for surface extraction.

Main Results:

• CAP-UDF demonstrates significant improvements in surface reconstruction from various point cloud sources (raw, scans, depth maps).
• The method achieves state-of-the-art performance compared to existing surface reconstruction techniques.
• Extended experiments show competitive performance in unsupervised point normal estimation.

Conclusions:

• CAP-UDF effectively addresses the limitations of previous UDF-based surface reconstruction methods.
• The proposed consistency-aware learning approach yields more accurate and smoother surface representations.
• CAP-UDF offers a robust and versatile solution for 3D computer vision tasks involving point cloud surface reconstruction.