Tensorflow implementation of Holoported Characters: Real-time Free-viewpoint Rendering of Humans from Sparse RGB Cameras.

CVPR, 2024

Ashwath Shetty^1,2, Marc Habermann^1,3, Guoxing Sun¹, Diogo Luvizon¹, Vladislav Golyanik^1,3, and Christian Theobalt^1,3

¹Max Planck Institute for Informatics,Saarland Informatics Campus , ²Saarland University , ³Saarbrücken Research Center for Visual Computing, Interaction and AI

This codebase contains the following VCAI/GVDH publications:

• Holoported Characters: Real-time Free-viewpoint Rendering of Humans from Sparse RGB Cameras (Shetty et al. 2024 (CVPR 2024)) Project Page

Further, it contains also more general tools such as a differentiable renderer, data readers, etc.

The code is structured in three main components:

1. The python-based deep learning framework (Projects). This contains the training code, texturing code, and general tools to process the Holported Characters Dataset
2. The custom Tensorflow operators, which are implemented in C++/CUDA (CustomTensorFlowCPPOperators)
3. The Cuda-based rasterizer (CudaRenderer)

You would need to install the custom Tensorflow and CUDA Rasterizer before if you would like to run training. It requires GCC 10.2, and we are working on a version to port it to modern setups.

Please note: This repo is a work in progress, please report any bugs or issues you face via email or issues, I would be happy to help out and improve it so that more people can use it. Please also create issues about features and information you would like

Anaconda / Python / Tensorflow Installation

• Create a conda environment: conda create --name tf280 python=3.9

• Activate your environment: conda activate tf280

• Install the cudatoolkit package: conda install -c nvidia cudatoolkit=11.2.0

• Install the cudatoolkit-dev package: conda install -c conda-forge cudatoolkit-dev=11.2.2

• Install tensorflow package: pip install tensorflow==2.8.0

• Tensorflow 2.8.0 does not natively finds the cuda package: To fix this do the following:

  • Open your bashrc_private: vi ~/.bashrc_private
  • Add the path to your cuda lib to the LD_LIBRARY_PATH variable.
  • Example: export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/HPS/RTMPC/work/Anaconda/envs/tf280/pkgs/cuda-toolkit/lib64
  • Close the script and run source ~/.bashrc_private

• Now test the tensorflow:

  • Login to the gpu20 server: srun -p gpu20 --gres gpu:1 --pty /bin/bash
  • Activate the new environment: conda activate tf280
  • Open python: python
  • Type: import tensorflow as tf
  • Type: t=tf.constant(8)
  • If no no-found errors are shown your conda/python/tensorflow is successfully installed If you are not using slurm, you can just run the code on your machine without logging into server Compiling the custom Tensorflow operators

• Go to CustomTensorFlowCPPOperators/cmakeTF2Linux

• Add the OpenCV dirs (Install OpenCV 3.4.1) for each partition, Tensorflow paths, and the nvcc path to your .bashrc_private. More precisely, define those variables: TENSORFLOW_PATH, OPENCV_PATH_RECON, OPENCV_PATH_GPU20 NVCC_PATH

• Therefore, open this file with vi ~/.bashrc_privateand add at the end of the file

  • For example:
  • export TENSORFLOW_PATH=/HPS/RTMPC/work/Programs/Anaconda3-4.1.1/envs/tf220/lib/python3.5/site-packages/tensorflow
  • export OPENCV_PATH_RECON=/HPS/RTMPC/work/MultiViewSupervisionNRTracking/Code/CustomTensorFlowCPPOperators/thirdParty/linux/opencv-3.4.7/build2
  • export OPENCV_PATH_GPU20=/HPS/RTMPC/work/MultiViewSupervisionNRTracking/Code/CustomTensorFlowCPPOperators/thirdParty/linux/opencv-3.4.7/build2
  • export NVCC_PATH=/HPS/RTMPC/work/Anaconda/envs/tf280/pkgs/cuda-toolkit/bin/nvcc

• Run srun -p gpu20 --gres gpu:1 -t 00:05:00 submitBuild.sh. Alternatively, you can also specifiy -p recon, if you do not use slurm bash submitBuild.sh

• Run srun -p gpu20 --gres gpu:1 -t 00:05:00 submitCompile.sh. Alternatively, you can also specifiy -p recon, if you do not use slurm bash submitCompile.sh

• If everything goes well you should find a file named libCustomTensorFlowOperators.so in CustomTensorFlowCPPOperators/binaries/Linux/ReleaseGpu20 or CustomTensorFlowCPPOperators/binaries/Linux/ReleaseRecon, respectively if not please create the folder and place the binaries in them (it should be in build directory)

Compiling the CUDA-based rasterizer

• Go to CudaRenderer/cpp/cmakeTF2Linux
• Run srun -p gpu20 --gres gpu:1 -t 00:05:00 submitBuild.sh. Alternatively, you can also specifiy -p recon,if you do not use slurm bash submitBuild.sh
• Run srun -p gpu20 --gres gpu:1 -t 00:05:00 submitCompile.sh. Alternatively, you can also specifiy -p recon, if you do not use slurm bash submitCompile.sh
• If everything goes well you should find a file named libCudaRenderer.so in CudaRenderer/cpp/binaries/Linux/ReleaseGpu20 or CudaRenderer/cpp/binaries/Linux/ReleaseRecon, respectively, respectively if not please create the folder and place the binaries in them (it should be in build directory)

Congratulations, you can now go to the Projects folder and run the python-based code. (This is the hardest part, everything else should go smoothly :))

For running a project please enter the Projects directory and follow the instructions. Note: Make sure that you activate your environment before trying to run one of the python scripts.

Both the source code and our data are under the terms of the Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. This project is only for research or education purposes, and not freely available for commercial use or redistribution.

@misc{shetty2023holoported,
      title={Holoported Characters: Real-time Free-viewpoint Rendering of Humans from Sparse RGB Cameras}, 
      author={Ashwath Shetty and Marc Habermann and Guoxing Sun and Diogo Luvizon and Vladislav Golyanik and Christian Theobalt},
      year={2023},
      eprint={2312.07423},
      archivePrefix={arXiv},
      primaryClass={cs.CV}

}