Visual Boundary-Guided Pseudo-Labeling for Weakly Supervised 3D Point Cloud Segmentation in Indoor Environments.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Accurate 3D point cloud segmentation is crucial for indoor scene understanding.
• Weakly supervised learning methods for point cloud segmentation face challenges with foreground-background imbalance due to complex indoor environments.
• Manual annotation of 3D point clouds is labor-intensive and time-consuming.

Purpose of the Study:

• To develop a novel foreground-aware label enhancement method for weakly supervised 3D point cloud segmentation.
• To address the performance imbalance between foreground and background elements in indoor scene segmentation.
• To introduce a versatile and portable module that improves existing weakly supervised segmentation techniques.

Main Methods:

• Projecting 3D point clouds onto 2D planes for 2D image segmentation and pseudo-label generation.
• Back-projecting 2D pseudo-labels into 3D space to train an initial segmentation model.
• Utilizing visual boundary priors from projected images to refine pseudo-labels and retrain the model, forming the Foreground Boundary Prior (FBP) module.

Main Results:

• The proposed Foreground Boundary Prior (FBP) method significantly improves 3D point cloud segmentation performance.
• Demonstrated effectiveness across different architectural backbones on the 2D-3D-Semantic dataset.
• Achieved enhanced segmentation accuracy using both random-sample and bounding-box based weak labeling strategies.

Conclusions:

• The Foreground Boundary Prior (FBP) is an effective plug-and-play module for enhancing weakly supervised point cloud segmentation.
• The method successfully mitigates foreground-background imbalance in indoor scene segmentation.
• The approach offers a portable solution for improving 3D point cloud analysis in various applications.