PUDM

This repo is the official project repository of the paper A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling.

• [ paper ] [ supp ]

The Overall Framework

Overview

• Citation
• Installation
• Data Preparation
• Model Zoo
• Quick Start

Citation

If you find PUDM useful to your research, please cite our work as an acknowledgment.

@InProceedings{Qu_2024_CVPR,
    author    = {Qu, Wentao and Shao, Yuantian and Meng, Lingwu and Huang, Xiaoshui and Xiao, Liang},
    title     = {A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {20786-20795}
}

Installation

Requirements

The following environment is recommended for running PUDM (an NVIDIA 3090 GPU):

• Ubuntu: 18.04 and above
• CUDA: 11.1 and above
• PyTorch: 1.9.1 and above
• python: 3.7 and above

Environment

• Base environment

conda create -n pudm python=3.7 -y
conda activate pudm

conda install cudatoolkit
pip install nvidia-cudnn-cu11

pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html

pip install open3d termcolor tqdm einops transforms3d==0.3.1
pip install msgpack-numpy lmdb h5py hydra-core==0.11.3 pytorch-lightning==0.7.1
pip install scikit-image black usort flake8 matplotlib jupyter imageio fvcore plotly opencv-python

# For installing pytorch3d, please follow:
1. download pytorch3d-0.6.1-py37_cu111_pyt191.tar.bz2 from https://anaconda.org/pytorch3d/pytorch3d/files?page=10
2. conda install pytorch3d-0.6.1-py37_cu111_pyt191.tar.bz2

# compile C++ extension packages
sh compile.sh

Data Preparation

Please download [ PU1K ] and [ PUGAN ].

# For generating test data, please see **PUDM-main/pointnet2/dataloder/prepare_dataset.py**
cd PUDM-main/pointnet2/dataloder

# For example 1, we can generate 4x test set of PUGAN:
python prepare_dataset.py --input_pts_num 2048 --R 4 --mesh_dir mesh_dir --save_dir save_dir

# For example 2, we can generate 4x test set of PUGAN with 0.1 Gaussion noise:
python prepare_dataset.py --input_pts_num 2048 --R 4 --noise_level 0.1 --noise_type gaussian --mesh_dir mesh_dir --save_dir save_dir

Model Zoo

Please download our checkpoints:
[ Baidu Netdisk ] (code : r2h9) or [ Google Drive ]
Please put checkpoints in the PUDM-main/pointnet2/pkls folder.
*Released model weights are temporarily as the model structure of PUDM may be adjusted later.

Quick Start

Example

We provide some examples. There examples are in the PUDM-main/pointnet2/example folder. The results are in the PUDM-main/pointnet2/test/example folder.

# For example, we can run 30 steps (DDIM) to generate 4x point cloud on KITTI with the pre-trained model of PUGAN.
# We provide the function (bin2xyz) of converting *.bin to *.xyz in **PUDM-main/pointnet2/dataloder/dataset_utils.py**.
cd PUDM-main/pointnet2
python example_samples.py --dataset PUGAN --R 4 --step 30 --example_file ./example/KITTI.xyz

This will produce the following result:

Training

We provide two datasets to train PUDM, PUGAN and PU1K. The results are in the PUDM-main/pointnet2/exp_{dataset} folder.

# For training PUGAN
cd PUDM-main/pointnet2
python train.py --dataset PUGAN

# For training PU1K
cd PUDM-main/pointnet2
python train.py --dataset PU1K

Testing

We provide two datasets to test PUDM. The results are in the PUDM-main/pointnet2/test/{dataset} folder.

# For testing PUGAN
cd PUDM-main/pointnet2
python samples.py --dataset PUGAN --R 4 --step 30 --batch_size 27

# For testing PU1K
cd PUDM-main/pointnet2
python samples.py --dataset PU1K --R 4 --step 30 --batch_size 43