(CVPR'24) Point Cloud Upsampling with Kernel Point Representation and Deformation

Installation

Step1. Install requirements:

python == 3.6.13
torch == 1.10.1
CUDA == 12.2
numpy == 1.19.5
open3d == 0.9.0.0
einops ==0.4.1
scikit-learn==1.0.1
tqdm==4.64.0
h5py==3.1.0

Step2. Compile the C++ extension modules:

cd models/Chamfer3D
python setup.py install
cd ../pointops
python setup.py install

Data preparation

Datasets can be download from here:

• We provide a pre-processed PU-GAN testing set with multiple resolutions of GT point clouds.
• If you want to generate testing point clouds from mesh files by youself, please refer to here.

After data preparation, the overall directory structure should be:

│data/
├──PU-GAN/
│   ├──train/
│   ├──test/
│   │   ├──pugan_4x
│   │   ├──pugan_16x
│   │   ├──arbitrary_scale
│   │   ├──.......
├──PU1K/
│   ├──train/
│   ├──test/

Training

Training models on PU-GAN (or PU1K) dataset:

python train.py --dataset pugan

or

python train.py --dataset pu1k

Results will be saved under ./output

Testing & Evaluation

We provide several pre-trained weights:

* indicates the origin model used in our paper

Testing example:

# 4X upsampling on PU-GAN dataset
python test.py --dataset pugan --input_dir ./data/PU-GAN/test/pugan_4x/input --gt_dir ./data/PU-GAN/test/pugan_4x/gt --ckpt ./pretrain/pugan_best.pth  --r 4 --save_dir ./result/pugan_4x

# 16X upsampling on PU-GAN dataset
python test.py --dataset pugan --input_dir ./data/PU-GAN/test/pugan_16x/input --gt_dir ./data/PU-GAN/test/pugan_16x/gt --ckpt ./pretrain/pugan_best.pth  --r 16 --save_dir ./result/pugan_16x

# 4X upsampling on PU1K dataset
python test.py --dataset pu1k --input_dir ./data/PU1K/test/input_2048/input_2048/ --gt_dir ./data/PU1K/test/input_2048/gt_8192/ --ckpt ./pretrain/pu1k_best.pth  --r 4 --save_dir ./result/pu1k_4x

# arbitrary-scale upsampling on PU-GAN dataset, take 19x for example
python test.py --dataset pugan --input_dir ./data/PU-GAN/test/arbitrary_scale/19x/input --gt_dir ./data/PU-GAN/test/arbitrary_scale/19x/gt --ckpt ./pretrain/pugan_best.pth  --r 19 --save_dir ./result/pugan_19x --flexible

• Don't miss "--flexible" for arbitrary-scale upsampling.
• If you want to use our original model, please use "--o", like:

  python test.py --dataset pugan --input_dir ./data/PU-GAN/test/pugan_4x/input --gt_dir ./data/PU-GAN/test/pugan_4x/gt --ckpt ./pretrain/pugan_o_best.pth  --r 4 --save_dir ./result/pugan_4x --o

• You can use our code to get CD value. To calculate HD and P2F value, please refer to here.

Surface reconstruction:

python surf_recon.py --file_path xxx.xyz --save_path xxx.obj

• Here, we provide an example point cloud and reconstructed result.

  Acknowledgements

  This repo is heavily based on KPConv, Grad-PU, PU-GCN, PU-GAN. Thanks for their great work!

Citation

@inproceedings{rong2024repkpu,
  title={RepKPU: Point Cloud Upsampling with Kernel Point Representation and Deformation},
  author={Rong, Yi and Zhou, Haoran and Xia, Kang and Mei, Cheng and Wang, Jiahao and Lu, Tong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={21050--21060},
  year={2024}
}