Self-Supervised Learning for Point-Cloud Classification by a Multigrid Autoencoder.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Direct processing of point clouds using shared multilayer perceptrons and aggregate functions is standard but struggles with local information capture, causing data loss.
• Existing methods establishing point-to-point relationships often do not fully resolve this information loss issue.

Purpose of the Study:

• To enhance the network's comprehension of point clouds by mitigating information loss inherent in current processing methods.
• To introduce a novel self-supervised approach for improving point cloud understanding and classification accuracy.

Main Methods:

• Proposed a multigrid autoencoder (MA) that integrates an autoencoder into the encoder of a classification network.
• Employed self-supervised learning to train the network, compelling the encoder to learn point cloud representations through reconstruction.
• Validated the MA model by integrating it with the PointNet++ architecture.

Main Results:

• The self-supervised multigrid autoencoder approach demonstrated improved performance over baseline methods.
• Achieved a 2.0% increase in classification accuracy on the ModelNet40 dataset.
• Observed a 4.7% improvement in classification accuracy on the ScanObjectNN dataset.

Conclusions:

• The proposed multigrid autoencoder effectively addresses information loss in point cloud processing.
• Self-supervised learning significantly enhances the performance of point cloud classification networks.
• The method offers a viable strategy for improving the understanding and analysis of 3D point cloud data.