HSGAN: Hierarchical Graph Learning for Point Cloud Generation.

Area of Science:

• Computer Vision
• 3D Shape Analysis
• Geometric Deep Learning

Background:

• Point clouds are versatile data representations for 3D objects, crucial in science and engineering.
• While point cloud learning excels in shape estimation and segmentation, unsupervised generation of 3D object parts remains challenging.
• Existing methods struggle with capturing complex topological and geometric structures in 3D shape generation.

Purpose of the Study:

• To propose a novel Generative Adversarial Network (GAN), Hierarchical Self-Attention GAN (HSGAN), for unsupervised 3D object part generation.
• To leverage graph convolutional networks (GCN) and self-attention mechanisms for enhanced 3D shape generation.
• To address limitations in existing generative pipelines, particularly in preserving geometric structures and training stability.

Main Methods:

• Developed HSGAN, a generative model that hierarchically transforms random codes into representation graphs using GCN and self-attention.
• Integrated global graph topology and latent topological information into the generative process.
• Introduced a new adversarial loss function to ensure training stability and prevent mode collapse.

Main Results:

• HSGAN effectively generates realistic 3D point clouds by utilizing compact latent topology as graph representations.
• The model avoids multiple discriminator updates per generator update, streamlining the training process.
• HSGAN successfully preserves dominant geometric structures throughout the hierarchical sampling process.

Conclusions:

• HSGAN offers a significant advancement in unsupervised 3D shape generation, producing realistic point clouds with preserved geometric integrity.
• The proposed architecture demonstrates robust performance and training stability, overcoming common GAN challenges like mode collapse.
• HSGAN shows potential as a versatile plug-and-play decoder within auto-encoding frameworks for 3D shape understanding.