Reproduce “Point Transformer” on OpenResearch

About this paper

Nico Engel, Vasileios Belagiannis, Klaus Dietmayer

A deep neural network called Point Transformer is introduced for 3D point cloud processing, which employs a novel SortNet module for learned point selection and ordering, combined with local-global attention mechanisms. This architecture achieves 92.8% accuracy on ModelNet40 for classification and 85.9% mIoU on ShapeNet for part segmentation, demonstrating robustness to rotations.

Key insights

• The learned scoring in SortNet effectively identifies geometrically 'important' points, empirically outperforming random sampling or furthest point sampling for feature selection.
• Combining ordered local features from SortNet with global features via cross-attention allows for a comprehensive representation that captures both fine details and overall shape without information loss from fixed-size pooling.
• The attention-based architecture demonstrates inherent robustness to input rotations, suggesting it learns fundamental geometric properties of objects.

Summary by alphaXiv.