T3DNet: Compressing Point Cloud Models for Lightweight 3-D Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• 3-D point cloud models are crucial for mobile applications like autonomous driving but are often too large for edge devices.
• Existing compression methods for 3-D point cloud models face challenges with memory and latency.
• There is a need for efficient methods to create lightweight 3-D point cloud models.

Purpose of the Study:

• To propose a novel method, T3DNet, for compressing large 3-D point cloud models into lightweight versions.
• To enable the deployment of 3-D point cloud models on resource-constrained edge devices.
• To achieve high compression rates with minimal accuracy loss.

Main Methods:

• T3DNet utilizes network augmentation and knowledge distillation to train a predefined tiny model.
• It improves performance through auxiliary supervision from augmented networks and the original model.
• The approach avoids traditional parameter reduction techniques like pruning or quantization.

Main Results:

• T3DNet achieves high compression rates on datasets like ModelNet40, ShapeNet, and ScanObjectNN.
• The method demonstrates state-of-the-art performance compared to existing compression techniques.
• On ModelNet40, T3DNet is 58% smaller and 54% faster with only a 1.4% accuracy decrease.

Conclusions:

• T3DNet offers an effective solution for creating lightweight 3-D point cloud models for mobile applications.
• The proposed method enables efficient deployment on edge devices without significant performance degradation.
• T3DNet represents a significant advancement in the compression of 3-D point cloud networks.