SSRNet: Scalable 3D Surface Reconstruction Network.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Learning-based surface reconstruction offers high expressiveness but struggles with scalability for large point clouds.
• Existing methods often fail to efficiently process extensive 3D datasets.

Purpose of the Study:

• To propose a novel, scalable learning-based 3D surface reconstruction method for large-scale point clouds.
• To address the limitations of existing methods in terms of processing efficiency and scalability.

Main Methods:

• Developed SSRNet, a method utilizing an octree structure for scalable reconstruction.
• Implemented a parallel processing pipeline that divides point clouds into local parts.
• Constructed local geometric features for octree vertices, capturing the relation to the implicit surface.

Main Results:

• SSRNet demonstrates state-of-the-art performance in 3D surface reconstruction.
• The method achieves outstanding efficiency in processing large-scale point clouds.
• A lightweight version of SSRNet offers a five-fold speed increase with maintained performance.

Conclusions:

• SSRNet provides a scalable and efficient solution for learning-based 3D surface reconstruction.
• The octree-based approach and local feature extraction prevent overfitting and ensure generalization.
• SSRNet significantly advances the state-of-the-art in processing large-scale 3D data.