PUFA-GAN: A Frequency-Aware Generative Adversarial Network for 3D Point Cloud Upsampling.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Point cloud upsampling is crucial for enhancing 3D data resolution.
• Existing methods struggle with uniform point distribution and capturing high-frequency details.

Purpose of the Study:

• To propose a novel generative adversarial network (GAN) for effective point cloud upsampling.
• To ensure evenly distributed points and generate clean high-frequency regions in upsampled point clouds.

Main Methods:

• Utilizing a GAN framework with a generator incorporating dynamic graph hierarchical residual aggregation for feature extraction and upsampling.
• Employing a discriminator with a graph filter to identify and retain high-frequency points for edge detail generation.
• Introducing an identity distribution loss function to maintain point distribution on the original surface.

Main Results:

• The proposed method generates upsampled point clouds with improved point distribution and detailed high-frequency regions.
• Objective and subjective evaluations show superior visual quality compared to state-of-the-art upsampling techniques.
• New metrics were introduced to specifically assess point regularity in high-frequency areas.

Conclusions:

• The developed GAN effectively addresses limitations in point cloud upsampling.
• The method achieves superior visual quality and point distribution, advancing the field of 3D data processing.