Awesome

Frequency-Modulated Point Cloud Rendering with Easy Editing

This repository contains the official implementation for the paper: Frequency-Modulated Point Cloud Rendering with Easy Editing (CVPR 2023 Highlight).

<img src="image/hotdog_chair.gif" width = "150" height = "150" alt="chair" /><img src="image/lego_chair.gif" width = "150" height = "150" alt="chair" /><img src="image/family_ficus.gif" width = "150" height = "150" alt="chair" /><img src="image/materials_drums.gif" width = "150" height = "150" alt="chair" />

Full code, configs and data will be available later.

Installation

conda create -n FreqPCR python=3.8
conda activate FreqPCR

conda install pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=10.2 -c pytorch
pip install matplotlib
pip install opencv-python
pip install lpips
pip install git+https://github.com/francois-rozet/piqa
pip install tensorboard
pip install ConfigArgParse
pip install open3d

# PyTorch3D rasterization
python setup.py develop

Optionally, for real-time rendering, please run the following command:

# OpenGL, borrowed from NPBG
pip install \
    Cython \
    PyOpenGL \
    PyOpenGL_accelerate

# need to install separately
pip install \
    git+https://github.com/DmitryUlyanov/glumpy \
    numpy-quaternion

# pycuda
git clone https://github.com/inducer/pycuda
cd pycuda
git submodule update --init
export PATH=$PATH:/usr/local/cuda/bin
./configure.py --cuda-enable-gl
python setup.py install
cd ..

Data Preparation

The layout should look like this

FreqPCR
├── data
    ├── nerf_synthetic
    ├── TanksAndTemple
    ├── dtu
    |   ├── dtu_110
    │   │   │── cams_1
    │   │   │── image
    │   │   │── mask
    │   │   │── pc.ply
    |   ├── dtu_114
    |   ├── dtu_118
    ├── scannet
    │   │   │──0000
    |   │   │   │──color_select
    |   │   │   │──pose_select
    |   │   │   |──intrinsic
    |   │   │   |──pc.ply
    │   │   │──0043
    │   │   │──0045

• NeRF-Synthetic: Please download the dataset provided by NeRF and put the unpacked files in data/nerf_synthetic. Since Point-NeRF provide an implementation of point cloud generation using MVSNet, we can run Point-NeRF and save the point clouds in data/pc/. You can also download raw point clouds from here.

• Tanks and Temples: Please download the dataset provided by NSVF and put the unpacked files in data/TanksAndTemple. To generate point clouds, run Point-NeRF and save the point clouds in data/pc/.

• DTU: Please download images and masks from IDR and camera parameters from PatchmatchNet. We use the point clouds provided by NPBG++.

• ScanNet: Please download data from ScanNet and run select_scan.py to select the frames. We use the provided depth map to generate point clouds. For scene0043_00, the frames after 1000 are ignored because the camera parameters are -inf.

Running

We experiment on a single NVIDIA GeForce RTX 3090 (24G), if your GPU memory is not enough, you can set the batch_size to 1 or reduce the train_size.

Point Cloud Geometry Optmization

python pc_opt.py --config=configs/pc_opt/nerf_hotdog.txt

The optimized point cloud will be saved in logs_pc_opt/

Rasterization

python run_rasterize.py --config=configs/render/nerf_hotdog.txt

The point index and depth buffers will be saved in data/fragments.

Training

python train.py --config=configs/render/nerf_hotdog.txt

The results will be saved in logs/. You can also run tensorboard to monitor training and testing.

Real-time Rendering

python inference_gl.py --config=configs/render/nerf_hotdog.txt

Editing

TODO

Acknowledgements

• This project builds on our previous work RadianceMapping.
• The code of rasterization borrows a lot from PyTorch3D and NPBG.

Citation

If you find our work useful in your research, please consider citing:

@article{zhang2023frequency,
  title={Frequency-Modulated Point Cloud Rendering with Easy Editing},
  author={Zhang, Yi and Huang, Xiaoyang and Ni, Bingbing and Li, Teng and Zhang, Wenjun},
  journal={arXiv preprint arXiv:2303.07596},
  year={2023}
}