APP-Net: Auxiliary-Point-Based Push and Pull Operations for Efficient Point Cloud Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Point cloud neural networks require efficient neighbor feature aggregation.
• Existing methods suffer from redundant computations and high memory usage due to repeated feature calculations.
• Complex local aggregators in prior work increase processing time.

Purpose of the Study:

• To propose a new local aggregator with linear complexity for point cloud analysis.
• To reduce computational costs and memory consumption in point cloud networks.
• To enhance the efficiency and performance of point cloud processing pipelines.

Main Methods:

• Introduced an auxiliary container for feature exchange between source points and aggregation centers, avoiding redundant computations.
• Developed a novel local aggregator named APP (Auxiliary container-based Point cloud aggregator).
• Integrated an online normal estimation module to provide geometric information for improved modeling.

Main Results:

• The proposed APP aggregator achieves linear complexity, significantly improving efficiency.
• APP-Net demonstrates reduced memory consumption and faster processing speeds compared to existing methods.
• Achieved comparable accuracies in classification and semantic segmentation tasks.

Conclusions:

• APP-Net offers a more efficient solution for point cloud analysis with lower memory footprint.
• The method enables high-throughput processing, exemplified by over 10,000 samples per second in classification.
• The proposed approach effectively balances efficiency and accuracy in point cloud deep learning.