ULD-Net: 3D unsupervised learning by dense similarity learning with equivariant-crop

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Deep learning methods for point cloud analysis often require extensive manual labeling.
• Unsupervised representation learning offers high label efficiency but faces challenges in extracting useful features from unlabeled 3D point clouds.

Purpose of the Study:

• To propose a novel unsupervised learning approach, ULD-Net, for effective point cloud analysis.
• To develop dense similarity learning for maximizing consistency across global-local views at instance and point levels.

Main Methods:

• ULD-Net employs an equivariant-crop (equiv-crop) module for dense similarity learning between global and local views.
• A simple Siamese network with stop-gradient operation prevents network collapse, avoiding complex designs like negative pairs or momentum encoders.
• Feature separability constraint is utilized for more representative embeddings.

Main Results:

• ULD-Net achieves state-of-the-art results among context-based unsupervised methods for point cloud analysis.
• It demonstrates comparable performance to supervised models in shape classification and segmentation tasks.
• Outperforms the STRL method by 1.1% overall accuracy on linear SVM classification and shows gains in fine-tuning settings.

Conclusions:

• ULD-Net provides an effective and simple unsupervised learning framework for 3D point cloud analysis.
• The method achieves superior performance compared to existing unsupervised approaches and is competitive with supervised methods.
• It offers a promising direction for label-efficient 3D data representation learning.