Full Point Encoding for Local Feature Aggregation in 3-D Point Clouds.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Geometry Processing

Background:

• Current point cloud processing methods often fail to explicitly model correlations between local and global features.
• Aggregation techniques in existing methods do not fully capture the intricate relationships within point cloud data.

Purpose of the Study:

• To propose a novel 'full point encoding' approach for 3D point cloud processing.
• To develop new architectures, Full Point Convolution (FuPConv) and Full Point Transformer (FPTransformer), that integrate this encoding.
• To enhance the modeling of internal correlations between local and global geometric features.

Main Methods:

• Introduced full point encoding applicable to both convolution and transformer architectures.
• Developed FuPConv and FPTransformer, which adaptively learn weights from local and global geometric connections via correlation functions.
• Proposed a shape-aware downsampling block considering local shape and global context.

Main Results:

• FuPConv and FPTransformer demonstrated strong generalization and high performance in point cloud tasks.
• Achieved state-of-the-art semantic segmentation results: 76.8% mIoU on S3DIS sixfold and 73.1% on S3DIS Area 5.
• Validated efficacy across semantic segmentation, object detection, classification, and normal estimation.

Conclusions:

• Full point encoding effectively models local and global geometric relationships and their internal correlations.
• The proposed FuPConv and FPTransformer architectures significantly advance point cloud processing capabilities.
• The methods show superior performance and generalization on benchmark datasets for diverse 3D tasks.