NeuralTPS: Learning Signed Distance Functions Without Priors From Single Sparse Point Clouds.

Area of Science:

• 3D Computer Vision
• Geometric Deep Learning
• Computational Geometry

Background:

• Surface reconstruction from point clouds is crucial in 3D computer vision.
• Current methods struggle with geometric variations and sparse data due to reliance on large-scale supervision and learned priors.
• Generalization to unseen geometric variations, especially from extremely sparse point clouds, remains a challenge.

Purpose of the Study:

• To develop a neural network for direct Signed Distance Function (SDF) inference from single sparse point clouds.
• To overcome limitations of existing methods by avoiding signed distance supervision, learned priors, and normal data.
• To improve the generalization ability and accuracy of surface reconstruction for sparse and unseen point clouds.

Main Methods:

• A novel neural network infers SDFs directly from sparse point clouds without prior knowledge or normal data.
• Surface parameterization and SDF inference are learned end-to-end.
• Parameterized surfaces are used as coarse samplers to generate supervision for a Thin Plate Splines (TPS) based network, enabling statistical SDF inference.

Main Results:

• The proposed method significantly enhances generalization ability and accuracy on unseen point clouds.
• Experimental results demonstrate superior performance compared to state-of-the-art methods on both synthetic and real-world sparse point cloud datasets.
• The approach effectively addresses the challenges posed by data sparsity in surface reconstruction.

Conclusions:

• The developed neural network offers a robust solution for surface reconstruction from sparse point clouds.
• The end-to-end learning of surface parameterization and SDF inference, coupled with a novel supervision mining strategy, proves effective.
• This work advances the state-of-the-art in handling sparse data for accurate and generalizable 3D surface reconstruction.