Arbitrary-Scale Point Cloud Upsampling With Saliency-Aware Implicit Surface Guidance.

Area of Science:

• Computer Vision and Graphics
• Geometric Deep Learning
• 3D Data Processing

Background:

• Point cloud upsampling is crucial for 3D data analysis but faces challenges with supervised learning and data acquisition.
• Existing methods struggle with continuous densification and accurate recovery of fine geometric structures due to the lack of explicit connectivity in point clouds.

Purpose of the Study:

• To develop a self-supervised point cloud upsampling model that overcomes limitations of supervised methods and improves geometric detail preservation.
• To enable fine-grained point densification and accurate recovery of both global shape and intricate structures in point clouds.

Main Methods:

• Proposed a saliency-aware implicit surface sampling approach for point cloud upsampling.
• Introduced a three-stage architecture: a pre-trained saliency guidance block, a saliency-aware enhancer, and an upsampler.
• Correlated implicit surface reconstruction with salient point identification for guided sampling and detail enhancement.

Main Results:

• The model significantly improves point cloud upsampling quality compared to state-of-the-art methods.
• Achieved accurate retention of global shape and meticulous structures in upsampled point clouds.
• Demonstrated flexibility and availability across diverse datasets, including synthetic and real-world captured shapes with varying scales and distributions.

Conclusions:

• The proposed saliency-aware implicit surface sampling method offers an effective self-supervised solution for point cloud upsampling.
• The method successfully addresses challenges in data acquisition and geometric structure recovery, providing high-quality upsampled point clouds.