---

I am looking for a machine learning engineer/ machine learning researcher position in EU. If you know any opening position, please do not hesitate to contact me! My linkedin page: Ziming Wang

---

SE(3)-bi-equivariant Transformers for Point Cloud Assembly

This is the official implement of the NeurIPS24 paper "SE(3)-bi-equivariant Transformers for Point Cloud Assembly".

BITR

BITR (SE(3)-bi-equivariant Transformers) is used to solve point cloud assembly problems, where the goal is to align two pieces of point clouds (may not be overlapped).

Overlapped point clouds	Non-overlapped point clouds

BITR has 3 nice properties:

1. BITR is SE(3)-bi-equivariant: its performance is not influence by the initial input position. This implies that it does not need to be iterative like ICP.
2. BITR is swap-equivariant: its result is consistent when the inputs are swapped.
3. BITR is swap-equivariant: its result is consistent when the inputs are scaled.

In addition, it does not assume that the inputs are overlapped.

Requirement

1. dgl (for graph processing)
2. escnn (for equivariant network definition)
3. torch
4. einops
5. torch_cluster, torch_scatter
6. open3d

Usage

Please see script.txt for evaluation on 7-scenes, ASL, ShapeNet (airplane) and BB (wine bottle) datasets.

Checkpoints and Datasets

1. Checkpoints can be downloaded from https://drive.google.com/file/d/13PWIcqmWm42w6tHlzPTMyoCY5nl0WJbZ/view?usp=sharing . They should be placed at the ./saved_model folder.
2. The processed dataset can be downloaded from https://drive.google.com/file/d/13PWIcqmWm42w6tHlzPTMyoCY5nl0WJbZ/view?usp=sharing. They should be placed at the ./Data folder.

Reference

If you find the code useful, please cite the following paper.

@inproceedings{wang2024se,
title={SE (3)-bi-equivariant Transformers for Point Cloud Assembly},
author={Wang, Ziming and J{\"o}rnsten, Rebecka},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS)},
year={2024},
}

If you have any question, comment or thought, you are welcome to contact me at [email protected]