SPU+: Dimension Folding for Semantic Point Cloud Upsampling.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• Semantic Point Cloud Upsampling (SPU) reconstructs dense 3D point clouds from sparse inputs.
• Conventional methods face a dimensional bottleneck with high-dimensional feature vectors.
• Effective feature representation is crucial for semantic tasks in SPU.

Purpose of the Study:

• To propose a novel SPU method, SPU+, that addresses the dimensional bottleneck.
• To introduce dimension folding as an alternative strategy for handling high-dimensional features.
• To achieve state-of-the-art performance in semantic point cloud upsampling.

Main Methods:

• SPU+ decomposes high-dimensional features into g-dimensional packages for interaction.
• A 3D Residual Graph Convolution Block (3D-RGCB) enables efficient 3D package convolutions.
• A scaling-and-shuffling strategy is developed for large-scale upsampling.

Main Results:

• Dimension folding proves effective in enhancing feature representation for SPU.
• SPU+ achieves state-of-the-art performance on publicly available datasets.
• Analysis of covering number demonstrates advantages of the 3D package representation.

Conclusions:

• SPU+ offers a significant advancement in Semantic Point Cloud Upsampling.
• Dimension folding and 3D package representation are key innovations.
• The proposed method demonstrates superior performance and efficiency.