Neural-IMLS: Self-Supervised Implicit Moving Least-Squares Network for Surface Reconstruction.

Area of Science:

• Computer Vision
• Geometric Deep Learning
• 3D Reconstruction

Background:

• Surface reconstruction from point clouds is difficult, especially with noisy real-world scans lacking normal information.
• Existing methods struggle with noise, missing data, and unoriented point clouds.

Purpose of the Study:

• To introduce Neural-IMLS, a novel self-supervised approach for robust surface reconstruction.
• To learn a noise-resistant signed distance function (SDF) directly from unoriented point clouds.

Main Methods:

• Utilizes a dual representation learning approach combining Multilayer Perceptron (MLP) and Implicit Moving Least Squares (IMLS).
• Employs a mutual regularization mechanism where MLP and IMLS enhance each other's SDF and normal estimations.
• Self-supervised learning framework directly processes raw, unoriented point clouds.

Main Results:

• Neural-IMLS successfully reconstructs faithful shapes from noisy and incomplete point clouds.
• The method demonstrates robustness on various synthetic and real-scan benchmarks.
• Achieves accurate geometric details and sharp features due to the synergistic MLP-IMLS interaction.

Conclusions:

• Neural-IMLS provides a powerful and noise-resistant solution for 3D surface reconstruction.
• The self-supervised, dual-representation learning approach effectively handles challenging input data.
• The learned SDF accurately approximates the underlying surface, even with significant noise and missing data.