DMRDenoise

[Project Page] [Paper] [Code]

This is an implementation of the DMRDenoise model based on our paper "Differentiable Manifold Reconstruction for Point Cloud Denoising", which will appear in ACM MM 2020.

📢 News

• [Aug 15, 2021] Please check out our latest work on point cloud denoising: Score-Based Point Cloud Denoising (ICCV 2021) [Paper] [Code].

Installation

Recommended Environment

The code has been tested with the following environment:

Package	Version
PyTorch	1.5.1
CUDA and NVCC Compiler	10.0
scikit-learn	0.23.1
h5py	2.10.0
PyTorch Lightning	0.7.6

Install via Conda

We provide a yaml file that allows instantaneous environment setup. Run the following command and it will create a conda environment named DMRDenoise with all required packages installed:

conda env create -f environment.yml
conda activate DMRDenoise

Next, compile the EMD operator. Note that this step requires CUDA NVCC compiler:

cd ./ops/emd
python setup.py install

Install Manually

You may also set up the environment manually. We provide the instruction as follows:

# Create a conda environment
conda create --name DMRDenoise python=3.6
conda activate DMRDenoise
# Install required packages
conda install -y pytorch=1.5.1 torchvision=0.6.1 cudatoolkit=9.2 -c pytorch
conda install -y scikit-learn=0.23.1
conda install -y -c conda-forge h5py=2.10.0 pytorch-lightning=0.7.6
# Compile the EMD operator
# [NOTICE] This step requires CUDA NVCC compiler
cd ./ops/emd; python setup.py install; cd ../../

Dataset

The training and testing datasets can be downloaded here. Extract the two zip files into the data folder to use them.

Usage

Denoise

We provide two pretrained models in the pretrained folder. You may use the models to denoise your point clouds:

python denoise.py --input <input_xyz> --output <output_xyz> --ckpt ./pretrained/supervised/epoch=153.ckpt

If you have downloaded the datasets, you may barely run python denoise.py to see a quick example.

Train

Before training the model on your own, please make sure the training dataset is prepared.

Supervised training:

python train.py

Unsupervised training:

python train.py --loss_ds None --loss_rec unsupervised

We provide TensorBoard logs of our pretrained models for reference. To view the logs, run:

tensorboard --logdir ./pretrained

Citation

Please cite our paper if you use this code in your own work:

@inproceedings{luo2020differentiable,
  title={Differentiable Manifold Reconstruction for Point Cloud Denoising},
  author={Luo, Shitong and Hu, Wei},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  year={2020},
  month={October}
}