SAT-NGP

[ArXiv]

Official implementation of SAT-NGP, as presented in our paper:

SAT-NGP : Unleashing Neural Graphics Primitives for Fast Relightable Transient-Free 3D reconstruction from Satellite Imagery (IGARSS 2024)
Camille Billouard ¹, Dawa Derksen ¹, Emmanuelle Sarrazin ¹ and Bruno Vallet ^2
1CNES, ²Univ Gustave Eiffel, ENSG, IGN, LASTIG, F-94160

Environment Setup

Tested configurations :

CPU/GPU	runs
AMD EPYC Milan 7713 / NVIDIA A100	✅
Intel Xeon E5-2698 / NVIDIA V100	❌
AMD EPYC 7742 / NVIDIA A100	✅
Intel Core i512400F / RTX 4060Ti	✅
AMD EPYC Milan 7713 / NVIDIA A40	✅

Create conda env

conda create -p satngp -y python=3.8
conda activate MY_PATH/satngp
python -m pip install --upgrade pip
python -m pip install setuptools==69.5.1 

Install lib

conda install anaconda::libtiff -y
conda install libnvjpeg-dev -c nvidia -y
conda install -c conda-forge ncurses -y
conda install gdal==3.4.1 libgdal -y
conda install -c anaconda git -y

Dependencies

Install PyTorch with CUDA (this repo has been tested with CUDA 11.7) :

pip install torch==2.0.1+cu117 torchvision==0.15.2+cu117 --extra-index-url https://download.pytorch.org/whl/cu117
conda install -c "nvidia/label/cuda-11.7.1" cuda-toolkit -y

Environment variables

cd satngp
ln -s lib lib64 # we want to avoid problems when compiling some packages
export LD_LIBRARY_PATH="$PWD/lib64:$LD_LIBRARY_PATH"
export VENV_LIB_PATH="$PWD/lib64/python3.8/site-packages/"
LDFLAGS="-L$PWD/lib"
export CUDA_PATH=$PWD
cd ..

Getting the repo

git clone https://github.com/Ellimac0/SAT-NGP.git
cd SAT-NGP/
# https://github.com/rusty1s/pytorch_scatter

Install and compile Pytorch Scatter

mkdir dep_ext
cd dep_ext
git clone --branch pytorch_1_11 https://github.com/rusty1s/pytorch_scatter/ 
cd pytorch_scatter
pip install . -vvv # -vvv is verbose for debugging during installation
cd ../..

Install requirements

cd SAT-NGP
pip install -r requirements.txt

Dataset

The data came from the DFC2019 dataset and the link below from SAT-NeRF

mkdir data
cd data
wget https://github.com/centreborelli/satnerf/releases/download/EarthVision2022/dataset.zip
unzip dataset.zip -d dataset

Train

# in SAT-NGP
# may take a few minutes the first time, as the backend is compiled at .cache/torch_extensions/py38_cu117/ 
bash scripts/run_sat_ngp.sh data/dataset JAX_XXX 60000 1024

Citation

Accepted to IEEE International Geoscience and Remote Sensing Symposium (IGARSS) 2024.

@misc{billouard2024satngp,
      title={SAT-NGP : Unleashing Neural Graphics Primitives for Fast Relightable Transient-Free 3D reconstruction from Satellite Imagery}, 
      author={Camille Billouard and Dawa Derksen and Emmanuelle Sarrazin and Bruno Vallet},
      year={2024},
      eprint={2403.18711},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgements

This work was performed using HPC resources from CNES Computing Center (DOI 10.24400/263303/CNES C3). The authors would like to thank the Johns Hopkins University Applied Physics Laboratory and IARPA for providing the data used in this study, and the IEEE GRSS Image Analysis and Data Fusion Technical Committee for organizing the Data Fusion Contest. A portion of this work was build on top of :

• Credits to Jiaxiang Tang for excellent work :

  @misc{torch-ngp,
  Author = {Jiaxiang Tang},
  Year = {2022},
  Note = {https://github.com/ashawkey/torch-ngp},
  Title = {Torch-ngp: a PyTorch implementation of instant-ngp}
  }
  

the contributors of :

• Pytorch_scatter

and authors of :

• SatNerf