PointNorm: Dual Normalization is All You Need for Point Cloud Analysis

Overview of the proposed DualNorm in our PointNorm framework.

• Sampling and Grouping
• Point Normalization: Normalize grouped points to sampled points
• Reverse Point Normalization: Normalize sampled points to grouped points

This addresses point cloud irregularity and facilitates learning for subsequent layers.

PointNorm's "push-and-pull" strategy for optimizing the point cloud density.

• When Δ > 1 (left column), PointNorm pushes points apart, which increases the standard deviation and reduces the point cloud density.
• When Δ < 1 (right column), PointNorm pulls points together, which reduces the standard deviation and increases the point cloud density.
• When Δ = 1 (middle column), no action is performed, and both the standard deviation and the point cloud density stay the same.

PointNorm for shape classification and part segmentation. Given an input point cloud, PointNorm embeds point features, uses sampling-grouping and DualNorm to normalize points, and employs Residual Blocks to leverage hierarchical features for accurate classification and segmentation.