Meta-PU: An Arbitrary-Scale Upsampling Network for Point Cloud.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Deep Learning

Background:

• Point cloud upsampling is crucial for 3D reconstruction quality.
• Current deep learning methods require separate models for each scale factor, leading to inefficiency.
• This necessitates a more flexible and unified approach for point cloud upsampling.

Purpose of the Study:

• To develop a single deep learning model capable of point cloud upsampling across arbitrary scale factors.
• To address the limitations of existing methods that treat each scale factor as an independent task.
• To enhance the efficiency and practicality of point cloud upsampling in real-world applications.

Main Methods:

• Proposed Meta-PU, a novel method for point cloud upsampling.
• Utilized a backbone network with residual graph convolution (RGC) blocks.
• Incorporated a meta-subnetwork for dynamic weight adjustment and a farthest sampling block for varied point sampling.

Main Results:

• Meta-PU successfully supports point cloud upsampling for arbitrary scale factors using a single model.
• Simultaneous training on multiple scales proved mutually beneficial.
• Meta-PU demonstrated superior performance compared to methods trained for specific scale factors.

Conclusions:

• Meta-PU offers a unified and efficient solution for point cloud upsampling.
• The meta-learning approach enables adaptability to various scale factors.
• This method advances the field of 3D reconstruction by improving scalability and performance.