Efficient Point Cloud Processing With High-Dimensional Positional Encoding and Non-Local MLPs.

Area of Science:

• Computer Vision
• Machine Learning
• Geometric Deep Learning

Background:

• Multi-Layer Perceptron (MLP) models are foundational in point cloud processing.
• Their complex architectures hinder understanding and application.
• Existing methods often focus on abstraction stages, with recent advances in refinement stages.

Purpose of the Study:

• To develop a novel two-stage abstraction and refinement (ABS-REF) view for modular feature extraction in point clouds.
• To introduce High-dimensional Positional Encoding (HPE) to enhance MLP-based point cloud processing.
• To create efficient and effective MLP networks for point cloud analysis.

Main Methods:

• Proposed a two-stage ABS-REF framework for point cloud feature extraction.
• Introduced High-dimensional Positional Encoding (HPE) module to incorporate intrinsic positional information.
• Replaced local MLP operations with efficient non-local MLPs combined with HPE for local representation.

Main Results:

• Developed HPENets, MLP networks adhering to the ABS-REF paradigm with an HPE-based refinement stage.
• HPENets demonstrated a strong balance between efficiency and effectiveness across seven datasets and four tasks.
• HPENet outperformed PointNeXt in accuracy metrics (mAcc, mIoU, Cls. mIoU) with significantly reduced FLOPs.

Conclusions:

• The ABS-REF view provides a clearer understanding of MLP model evolution in point cloud processing.
• HPE module effectively integrates positional information into MLP architectures.
• HPENets offer a promising, efficient, and effective solution for various point cloud processing tasks.