Generative Variational-Contrastive Learning for Self-Supervised Point Cloud Representation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Self-supervised representation learning for 3D point clouds is crucial.
• Existing methods use fixed embeddings, limiting generalization across data domains.

Purpose of the Study:

• To propose a novel Generative Variational-Contrastive Learning (GVC) model.
• To improve the transferability of feature extractors between synthetic and real-world data.

Main Methods:

• Utilized Gaussian distribution for continuous, smoothed latent feature representation.
• Introduced a variational contrastive module to constrain feature distributions.
• Implemented a generative cross-supervision module for feature invariance and distribution consistency.

Main Results:

• GVC achieved state-of-the-art performance on various downstream tasks.
• Pre-training on synthetic data with GVC led to significant gains (8.4% and 14.2%) on real-world datasets for linear and few-shot classification.

Conclusions:

• GVC effectively addresses limitations of fixed embeddings in 3D point cloud learning.
• The proposed model demonstrates superior transfer learning capabilities for 3D computer vision tasks.