Rethinking Masked Representation Learning for 3D Point Cloud Understanding.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Self-supervised representation learning for point clouds is crucial for understanding unlabeled 3D data.
• Existing masked point modeling methods use overlapping groups, leading to information leakage and inconsistent feature representations for semantically similar parts.

Purpose of the Study:

• To develop a novel hierarchical masked representation learning method for point clouds.
• To address limitations of current grouping strategies and semantic modeling in self-supervised point cloud learning.

Main Methods:

• Proposed an optimal transport-based hierarchical grouping strategy for non-overlapping point cloud partitioning.
• Introduced a prototype-based part modeling module for consistent semantic feature representation.
• Utilized a hierarchical attention encoder for enhanced feature extraction.

Main Results:

• The proposed non-overlapping grouping strategy prevents early leakage of structural information.
• The prototype-based module ensures feature consistency for semantically similar object components.
• Achieved state-of-the-art performance on four downstream tasks, surpassing existing 3D representation learning methods.

Conclusions:

• The novel hierarchical masked representation learning method significantly enhances self-supervised point cloud understanding.
• The proposed modules effectively improve feature representation by addressing grouping and semantic modeling challenges.
• Experimental results and ablation studies validate the method's effectiveness and superiority.