Robust Point Cloud Segmentation With Noisy Annotations.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• 3D point cloud segmentation is crucial for various applications.
• Deep learning methods struggle with noisy labels common in real-world datasets.
• Existing noise-robust methods are insufficient for the unique challenges of point clouds.

Purpose of the Study:

• To develop a novel framework for robust point cloud segmentation under noisy labels.
• To address both instance-level and boundary-level label noise.
• To create a noise-rate blind approach adaptable to spatially variant noise.

Main Methods:

• Proposed a Point Noise-Adaptive Learning (PNAL) framework.
• Implemented point-wise confidence selection for reliable label extraction.
• Utilized cluster-wise label correction with voting for neighbor correlation.
• Introduced PNAL-boundary variant for progressive boundary label cleaning.

Main Results:

• PNAL significantly outperforms baseline methods on noisy datasets.
• The framework demonstrates effectiveness even with 60% symmetric noise and high boundary noise.
• Performance is comparable to models trained on completely clean data.
• Validated on synthetic and real-world datasets, including cleaned ScanNetV2.

Conclusions:

• PNAL offers a robust solution for 3D point cloud segmentation with noisy labels.
• The noise-rate blind and spatially adaptive nature makes it suitable for real-world data.
• PNAL-boundary effectively tackles boundary-level noise, enhancing segmentation accuracy.