Point-NAS: A Novel Neural Architecture Search Framework for Point Cloud Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• 3D point cloud processing is crucial for many applications.
• Designing effective neural networks for point cloud tasks is complex and time-consuming.
• Existing methods require extensive manual tuning of network parameters and hyperparameters.

Purpose of the Study:

• To develop an automated framework for discovering optimal neural network architectures for 3D point cloud tasks.
• To reduce the computational cost and manual effort associated with network design.
• To create a versatile architecture search method applicable to diverse point cloud challenges.

Main Methods:

• Introduced Point-NAS, a one-shot neural architecture search (NAS) framework.
• Designed an elastic feature extraction (EFE) module as a scalable building block.
• Developed a supernet encompassing diverse network structures and employed a weight coupling sandwich rule for optimization.
• Implemented a united gradient adjustment algorithm to enhance convergence and training stability.

Main Results:

• The trained supernet effectively optimizes numerous subnets for specific tasks.
• Achieved state-of-the-art results on benchmarks like ModelNet40 (94.2% accuracy) and ScanObjectNN (88.9% accuracy).
• Demonstrated strong performance in semantic segmentation (S3DIS, 68.6% mIoU) and object detection (SUN RGB-D, 63.6% mAP@0.25).

Conclusions:

• Point-NAS successfully automates the architecture search for 3D point cloud networks.
• The proposed framework offers an efficient and effective solution for optimizing network performance across various tasks.
• Inherited weights from the supernet enable high-performing, task-specific subnets without extensive retraining.