SPU-Net: Self-Supervised Point Cloud Upsampling by Coarse-to-Fine Reconstruction With Self-Projection Optimization.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Point cloud upsampling aims to increase point density and uniformity from sparse, irregular data.
• Current deep learning methods often require paired sparse-dense ground truth data for supervision.
• Acquiring real-world paired data is costly and time-consuming, limiting model applicability.

Purpose of the Study:

• To develop a self-supervised point cloud upsampling network (SPU-Net).
• To enable effective upsampling using inherent data patterns, eliminating the need for ground truth.
• To address limitations of supervised methods with real-scanned sparse data.

Main Methods:

• A coarse-to-fine reconstruction framework comprising point feature extraction and expansion.
• Integration of self-attention with Graph Convolutional Networks (GCN) for context aggregation.
• A hierarchical folding strategy using learnable 2D grids for point generation.
• A novel self-projection optimization with joint loss (uniformity and reconstruction) for refinement.

Main Results:

• SPU-Net successfully performs self-supervised point cloud upsampling.
• The method achieves performance comparable to state-of-the-art supervised approaches.
• Experiments on both synthetic and real-scanned datasets validate the effectiveness.

Conclusions:

• SPU-Net offers a viable self-supervised alternative for point cloud upsampling.
• The proposed framework effectively captures inherent upsampling patterns from sparse data.
• This approach reduces reliance on expensive ground truth data for real-world applications.