Point Cloud Deep Learning Network Based on Balanced Sampling and Hybrid Pooling.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Automatic semantic segmentation of point cloud data is crucial for machine vision, virtual reality, and smart cities.
• Existing methods, like PointNet++, struggle with extremely imbalanced point cloud scenes, necessitating improved processing capabilities.

Purpose of the Study:

• To enhance the processing capability of point cloud segmentation methods for imbalanced scenes.
• To improve the accuracy and efficiency of semantic segmentation in 3D data.

Main Methods:

• Designed a weighted sampling method based on farthest point sampling (FPS) that adjusts sampling weights using model loss to equalize sampling.
• Introduced relational learning within the neighborhood space using a self-attention model to distinguish feature importance during feature encoding.
• Employed a hybrid pooling method for aggregating and transmitting global-local features.

Main Results:

• Achieved 9.5% and 11.6% improvement in overall accuracy (OA) and mean of class-wise intersection over union (MIoU) on the S3DIS dataset compared to the baseline.
• Demonstrated 4.2% and 3.9% improvement in OA and MIoU on the Vaihingen dataset compared to the baseline.
• The proposed algorithm shows a favorable balance between OA and MIoU compared to other network models on public datasets.

Conclusions:

• The proposed network effectively addresses the challenge of imbalanced point cloud scenes in semantic segmentation.
• The integration of weighted sampling, relational learning, and hybrid pooling significantly enhances segmentation performance.
• This work offers a robust solution for accurate and balanced semantic segmentation of 3D point cloud data.