Rethink Rotation

Official implementation of "Rethinking Rotation Invariance with Point Cloud Registration", AAAI 2023

[Paper] [Supp.] [Video]

Requirements

To build the CUDA kernel for FPS:

pip install pointnet2_ops_lib/.  

NOTE: If you encounter problems while building the kernel, you can refer to Pointnet2_PyTorch for solutions.

Code

This repo contains Pytorch implementation of the following modules:

• [x] ModelNet40 Classification under rotations

  bash scripts/modelnet_cls.sh
  

• [x] ScanObjectNN Classification under rotations

  bash scripts/scanobject_cls.sh
  

• [ ] ShapeNetPart Segmentation under rotations

Performance

• State-of-the-art accuracy on ModelNet40 under rotation: 91.0% (z/z), 91.0% (z/SO(3)).
• State-of-the-art accuracy on ScanObjectNN OBJ_BG classification under rotation: 86.6% (z/z), 86.3% (z/SO(3)).
• State-of-the-art micro and macro mAP on ShapeNetCore55 under rotation: 0.715, 0.510.
• ShapeNetPart segmentation under rotation: 80.3% (z/z), 80.4% (z/SO(3)).

Citation

If you find this repo useful in your work or research, please cite:

Acknowledgement

Our code borrows a lot from:

• DGCNN
• DGCNN.pytorch
• PointContrast