Self-Supervised Arbitrary-Scale Implicit Point Clouds Upsampling.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Point clouds upsampling (PCU) is crucial for dense 3D data generation from sparse sensor inputs like LiDAR.
• Existing deep learning methods for PCU often require extensive paired data for supervised training or suffer from complexity due to scale-specific networks.

Purpose of the Study:

• To develop a self-supervised and magnification-flexible method for point clouds upsampling.
• To address the limitations of supervised training and complex multi-scale networks in PCU.

Main Methods:

• Formulated PCU as finding nearest projection points on an implicit surface for seed points.
• Defined two implicit neural functions for projection direction and distance estimation, trained via pretext tasks.
• Implemented a projection rectification strategy to remove outliers and preserve object sharpness.

Main Results:

• The proposed self-supervised learning scheme achieves competitive or superior performance compared to state-of-the-art supervised methods.
• Demonstrated the effectiveness of the implicit neural function approach for PCU.
• Validated the method's ability to generate dense and uniform point clouds without explicit ground truth.

Conclusions:

• The novel self-supervised PCU method offers a viable alternative to data-hungry supervised approaches.
• The magnification-flexible nature and improved performance highlight the potential of this technique for real-world applications.
• This approach simplifies PCU by avoiding the need for multiple networks for different scaling factors.