High-Quality Pseudo-Labeling for Point Cloud Segmentation With Scene-Level Annotation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Indoor point cloud semantic segmentation is crucial for understanding 3D environments.
• Scene-level annotation presents a significant challenge due to the lack of precise point-level labels.
• Existing methods struggle with generating accurate point-level pseudo-labels from scene-level data.

Purpose of the Study:

• To propose a high-quality pseudo-label generation framework for indoor point cloud semantic segmentation using scene-level annotations.
• To enhance segmentation accuracy by addressing the limitations of current pseudo-labeling techniques.
• To improve feature learning and semantic prediction in point cloud data.

Main Methods:

• A cross-modal feature guidance module utilizing 2D-3D correspondences to align point cloud and image features.
• A region-point semantic consistency module employing a region-voting strategy to guide point-level predictions.
• Leveraging multi-modal information and semantic consistency for accurate pseudo-label generation.

Main Results:

• Achieved significant improvements over previous methods on ScanNet v2 and S3DIS datasets under scene-level annotation.
• Demonstrated the effectiveness of the proposed framework in rectifying inaccurate point-level semantic predictions.
• Ablation studies validated the contribution of individual components of the approach.

Conclusions:

• The proposed framework effectively generates high-quality pseudo-labels for indoor point cloud semantic segmentation with scene-level annotations.
• The integration of cross-modal features and region-point semantic consistency significantly enhances segmentation performance.
• This work offers a promising direction for leveraging less granular annotations in 3D semantic segmentation tasks.