Generating Inverse Feature Space for Class Imbalance in Point Cloud Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Point cloud semantic segmentation is vital for understanding production environments.
• Deep learning model performance relies heavily on training data quality and balance.
• Imbalanced datasets can lead to cognitive bias in segmentation networks.

Purpose of the Study:

• To propose InvSpaceNet, a novel framework to mitigate cognitive bias caused by imbalanced data in point cloud semantic segmentation.
• To enhance the efficacy and generalization of deep learning-based segmentation models.

Main Methods:

• A dual-branch training architecture combining instance-balanced and inverse sampling data.
• Generation of an inverse feature space with contrastive loss for aggregated class points.
• Utilizing momentum-updated class prototypes from the inverse space to guide main branch segmentation.

Main Results:

• Demonstrated effective mitigation of point cloud data imbalance issues.
• Achieved improved segmentation performance across four large benchmarks (S3DIS, ScanNet v2, Toronto-3D, SemanticKITTI).

Conclusions:

• InvSpaceNet successfully alleviates cognitive bias stemming from imbalanced datasets.
• The proposed method offers a significant advancement in point cloud semantic segmentation accuracy and robustness.