Revolutionizing Point Cloud Classification with Multiview Self-Supervised Learning

Self-supervised models are shaping the future of point cloud processing by minimizing reliance on labeled data and addressing the challenges associated with point cloud annotation. Traditional self-supervised point cloud models often employ mask reconstruction, which involves globally dividing patches. This approach can lead to increased computational costs and insufficient consideration of global features. In this paper, we introduce MVMAE, a multiview-based self-supervised model designed for point cloud processing. MVMAE firstly converts point clouds into multiview images through rendering. It then integrates views from various angles as inputs to a mask self-encoder, where each view is treated as a distinct patch. This method provides a more comprehensive perspective for feature learning, enhancing point cloud classification performance. Experiments conducted on datasets like ModelNet40 demonstrate that MVMAE achieves a classification accuracy of 90.1% and offers improved processing speed while managing computational complexity and parameter count. The code for MVMAE is available at https://github.com/NWUzhouwei/MVMAE.