Point cloud generation adversarial network based on self-attention and curvature.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Point clouds are crucial for 3D data tasks like segmentation and classification.
• Existing point cloud generation models struggle with noise and uneven point distribution.
• Authentic generation of local geometric details remains a challenge.

Purpose of the Study:

• To propose a novel adversarial network, SAC-GAN, for high-quality point cloud generation.
• To enhance the authenticity and consistency of generated point clouds.
• To improve the discriminator's ability to extract local and global features.

Main Methods:

• Developed a generator with feature enhancement and pre-processing modules using the ShapeNetCore dataset.
• Adjusted the discriminator's loss function with Wasserstein distance and normal vectors.
• Integrated a self-attention mechanism into the discriminator for improved feature extraction.

Main Results:

• SAC-GAN demonstrated superior performance compared to TreeGAN, SP-GAN, PDGN, and WarpingGAN.
• Achieved a 4.24% reduction in Jensen-Shannon Divergence (JSD).
• Reduced Maximum Mean Discrepancy (MMD) by 0.8 and increased Coverage (COV) by 1.25%.

Conclusions:

• The proposed SAC-GAN model effectively generates point clouds with high shape integrity and authenticity.
• The integration of self-attention and curvature learning mechanisms significantly improves generation quality.
• SAC-GAN offers a robust solution for generating realistic 3D point cloud data.