.png){kind=icon}
.png){kind=icon}
Return
Weak label selection optimization framework for weakly supervised point cloud segmentation
.png){kind=icon}
PRE
AI
C
W
H
B
DOI：10.1016/j.displa.2026.103454.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Point cloud segmentation plays a crucial role in autonomous driving, robotic navigation, and medical image analysis, with significant practical applications. However, the heavy reliance on annotated data poses substantial challenges for large-scale deployment, making weakly supervised learning an increasingly important research direction. While numerous weakly supervised point cloud segmentation methods have been proposed, the critical issue of weak label initialization has been largely overlooked. Existing approaches predominantly rely on random selection strategies, resulting in label redundancy and suboptimal information utilization. In this paper, we propose an optimized weak label selection framework with the following key contributions: (1) We propose an unsupervised feature-based label selection framework that employs self-supervised learning to extract discriminative features, followed by density estimation and clustering to identify information-rich regions in the feature space. (2) We introduce a spatial-similarity constrained label deduplication module that enhances sampling diversity through joint global similarity and local spatial constraints. (3) We develop a cross-sample collaborative sampling strategy that constructs a unified feature space for semantically similar samples, leveraging fusion clustering to optimize sampling weights for improved information coverage and complementarity. Comprehensive experiments on ScanNet-v2 and S3DIS datasets demonstrate that our framework consistently outperforms both random selection and class-balanced selection baselines in weakly supervised point cloud segmentation tasks, achieving improvements of up to 2.01% in MIOU with only 0.1% labeled data on S3DIS dataset.
Keywords:
Point clouds
Semantic segmentation
Weakly supervised learning
Unsupervised learning
Label optimization
Journal
IF:
3.4
Papers:
2.1K
Citations:
3.2K
Organization
Cited Papers