Weakly Supervised Segmentation on Outdoor 4D Point Clouds With Progressive 4D Grouping.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Weakly supervised 3D point cloud segmentation methods aim to reduce annotation efforts.
• Previous work (W4DTS) used 0.001% of points for segmentation but suffered from low-quality pseudo-labels.
• Extremely limited annotations pose a challenge for effective pseudo-label generation.

Purpose of the Study:

• To develop a method for generating high-quality pseudo-labels with very sparse annotations in 3D point cloud segmentation.
• To improve the performance of weakly supervised segmentation models under extreme annotation constraints.
• To enhance the progressive 4D grouping approach with additional learning strategies.

Main Methods:

• Proposed a progressive 4D grouping approach to aggregate spatial and temporal information from sparse annotated and unannotated points.
• Introduced cross-frame contrastive learning to refine feature representations across different time frames.
• Implemented local consistency learning to enforce neighborhood agreement in the segmentation process.

Main Results:

• Achieved significant performance improvements on the SemanticKITTI dataset using only 0.001% annotations, outperforming the previous best approach.
• Demonstrated competitive results on SemanticPOSS and ScribbleKITTI datasets, nearing the performance of fully supervised models.
• Validated the effectiveness of the progressive 4D grouping, cross-frame contrastive learning, and local consistency learning.

Conclusions:

• The proposed progressive 4D grouping approach effectively generates high-quality pseudo-labels even with extremely sparse annotations.
• This method significantly advances weakly supervised 3D point cloud segmentation under severe annotation budget limitations.
• The framework offers a promising direction for efficient 3D data annotation and segmentation in real-world applications.