PointAttN

News：Our paper has been accepted by AAAI-2024.

1. Environment setup

Install related libraries

This code has been tested on Ubuntu 20.04, python 3.8.12, torch 1.9.0 and cuda 11.2. Please install related libraries before running this code:

pip install -r requirements.txt

Compile Pytorch 3rd-party modules

please compile Pytorch 3rd-party modules ChamferDistancePytorch and mm3d_pn2. A simple way is using the following command:

cd $PointAttN_Home/utils/ChamferDistancePytorch/chamfer3D
python setup.py install

cd $PointAttN_Home/utils/mm3d_pn2
python setup.py build_ext --inplace

2. Train

Prepare training datasets

Download the datasets:

Train a model

To train the PointAttN model, modify the dataset path in cfgs/PointAttN.yaml , run:

python train.py -c PointAttN.yaml

3. Test

Pretrained models

The pretrained models on Completion3D and PCN benchmark are available as follows:

dataset	performance	model link
Completion3D	CD = 6.63	[BaiDuYun] (code：nf0m)[GoogleDrive]
PCN	CD = 6.86	[BaiDuYun] (code：kmju)[GoogleDrive]

Test for paper result

To test PointAttN on PCN benchmark, download the pretrained model and put it into PointAttN_cd_debug_pcn directory, run:

python test_pcn.py -c PointAttN.yaml

To test PointAttN on Completion3D benchmark, download the pretrained model and put it into PointAttN_cd_debug_c3d directory, run:

python test_c3d.py -c PointAttN.yaml

4. Acknowledgement

We include the following PyTorch 3rd-party libraries:
[1] ChamferDistancePytorch
[2] mm3d_pn2
Some of the code of this project is borrowed from VRC-Net

5. Cite this work

If you use PointAttN in your work, please cite our paper:

@InProceedings{Wang_2024_AAAI,
   author = {Jun Wang, Ying Cui, Dongyan Guo, Junxia Li, Qingshan Liu, Chunhua Shen},
   title = {PointAttN: You Only Need Attention for Point Cloud Completion},
   booktitle = {Association for the Advancement of Artificial Intelligence (AAAI)},
   month = {Feb},
   year = {2024}
 }

hanbinbin-web/PointAttN