Here is the official repo for the paper:

• FoV-NeRF: Foveated Neural Radiance Field for Virtual Reality (Deng et al., 2022, TVCG (ISMAR 2022)).

Virtual Reality (VR) is becoming ubiquitous with the rise of consumer displays and commercial VR platforms. Such displays require low latency and high quality rendering of synthetic imagery with reduced compute overheads. Recent advances in neural rendering showed promise of unlocking new possibilities in 3D computer graphics via image-based representations of virtual or physical environments. Specifically, the neural radiance fields (NeRF) demonstrated that photo-realistic quality and continuous view changes of 3D scenes can be achieved without loss of view-dependent effects. While NeRF can significantly benefit rendering for VR applications, it faces unique challenges posed by high field-of-view, high resolution, and stereoscopic/egocentric viewing, typically causing low quality and high latency of the rendered images. In VR, this not only harms the interaction experience but may also cause sickness.

To tackle these problems toward six-degrees-of-freedom, egocentric, and stereo NeRF in VR, we present the first gaze-contingent 3D neural representation and view synthesis method. We incorporate the human psychophysics of visual- and stereo-acuity into an egocentric neural representation of 3D scenery. We then jointly optimize the latency/performance and visual quality while mutually bridging human perception and neural scene synthesis to achieve perceptually high-quality immersive interaction. We conducted both objective analysis and subjective studies to evaluate the effectiveness of our approach. We find that our method significantly reduces latency (up to 99% time reduction compared with NeRF) without loss of high-fidelity rendering (perceptually identical to full-resolution ground truth). The presented approach may serve as the first step toward future VR/AR systems that capture, teleport, and visualize remote environments in real-time.

• Installation
• Dataset
• Usage
  • Training
  • Evaluation
  • Foveated Neural Rendering
  • Real-time Interactive Rendering
• Q&A
• Citation

This code is implemented in PyTorch and C++. The code has been tested on the following system:

• Python 3.10
• PyTorch 1.11.0
• CUDA 11.3
• TensorRT 8.0
• Ubuntu 18.04
• Nvidia GeForce RTX 3090

Install all dependencies by:

conda create -n fov_nerf python=3.10
conda activate fov_nerf
conda install ffmpeg -c conda-forge
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
conda install numpy tqdm matplotlib tensorboard configargparse
pip install PyGLM tensorboardX lpips

Install dependencies first, then compile the program from the source code.

Note: Commands below assume you have set the environment variable ${CUDA_INSTALL_DIR} to your cuda install directory (e.g. /usr/local/cuda)

Install cuDNN

1. Download the latest cuDNN tar package.
2. Extract the downloaded package to some directory.
3. Copy all contents in include/, lib/ to corresponding folders in CUDA install directory:

cd ${CUDNN_EXTRACT_DIR} && \
sudo cp -r include/* ${CUDA_INSTALL_DIR}/include/ && \
sudo cp -r lib/* ${CUDA_INSTALL_DIR}/lib64/ && \

Install TensorRT

1. Download TensorRT 8.0 tar package.
2. Extract the downloaded package to some directory.
3. Build trtexec:

cd ${TENSORRT_EXTRACT_DIR}/samples/trtexec && make

4. Copy all contents in include/, lib/ and trtexec to corresponding folders in CUDA install directory:

cd ${TENSORRT_EXTRACT_DIR} && \
sudo cp -r include/* ${CUDA_INSTALL_DIR}/include/ && \
sudo cp -r lib/* ${CUDA_INSTALL_DIR}/lib64/ && \
sudo cp bin/trtexec ${CUDA_INSTALL_DIR}/bin/

5. Run trtexec to validate the installation.

If you got some error like "libnvinfer.so.8 not found", add ${CUDA_INSTALL_DIR}/lib64 to the LD_LIBRARY_PATH environment variable and retry:

touch ~/.bashrc && \
echo 'export LD_LIBRARY_PATH=${CUDA_INSTALL_DIR}/lib64:${LD_LIBRARY_PATH}' >> ~/.bashrc && \
source ~/.bashrc

Install glfw, GLEW and GLM

sudo apt install libglfw3 libglfw3-dev libglew2.1 libglew-dev glew-utils libglm-dev

Compile

cd cpp/view_gl && make

When successfully compiled, the executable file will be generated under cpp/bin folder.

You can download the datasets used in our paper from Google Drive.

Please also cite the original papers if you use any of them in your work.

To prepare a new dataset of a single scene for training and testing, please follow the data structure:

${SCENE}
|-- ${TRAIN_DATASET}.json     # dataset description file
|-- ${TRAIN_DATASET}
    |-- view_0000.png         # target image for each view
    |-- view_0001.png
    ...
|-- ${TEST_DATASET}.json      # dataset description file
|-- ${TEST_DATASET}
    |-- view_0000.png         # target image for each view
    |-- view_0001.png
    ...

The ${*_DATASET}.json is a JSON format file containing camera intrinsic as well as the poses of every views in ${*_DATASET}/. Please check our dataset for an example.

Use following command to start a new training:

python train.py -c ${CONFIG} --trainset ${SCENE_FOLDER}/${TRAIN_DATASET}.json

The ${CONFIG} is the name of a config file in configs/ without suffix. The logs and checkpoints will be stored in ${SCENE_FOLDER}/_nets/${TRAIN_DATASET}/${CONFIG}

For example, the following command will train a foveal model for the example scene barbershop:

python train.py -c fovea --trainset data/barbershop/fovea_train.json

To override some of the arguments in config file for current training, simply append the arguments to the command. For example, the following command will train a model with 8 layers instead of 4 and up to 10 epochs instead of 200 epochs:

python train.py -c fovea --trainset data/barbershop/fovea_train.json --depth 8 --max-epochs 10

Please refer to example config files in configs/ for the complete set of arguments.

You can also continue training on an existing checkpoint by:

python train.py -p ${CHECKPOINT_FILE}

Note: The arguments for trainer (those argument below the line trainer=... in config files) can be overridden by command-line arguments in continue-training mode.

Once the model is trained, the following command is used to evaluate rendering quality on the test dataset:

python test.py ${CHECKPOINT_PATH} ${SCENE_FOLDER}/${TEST_DATASET}.json

The outputs will be stored in ${SCENE_FOLDER}/_nets/${TRAIN_DATASET}/${CONFIG}/output_*/${TEST_DATASET}_*/.

You can also use the script to render stereo views by append the argument --stereo ${IPD}, where ${IPD} is the inter-pupil distance in meters (generally set to 0.06). This will disable the metrics evalution.

The --media video argument will tell the script to generate a video instead of a sequence of images. This is useful for the dataset containing a set of continous moving views.

When both foveal and periphery models are trained for a scene, you can use the following command to render the foveated views from an existing view and gaze sequence:

python render.py ${VIEW_FILE} -m ${MODEL_FOLDER} [-f ${FPS}]

${VIEW_FILE} can be a JSON-format dataset description file containing poses and gazes of every view or a CSV file, each line of which contains the gaze and pose of a view (16 columns):

<left_gaze_in_pixels>, <right_gaze_in_pixels>, <translation_vector>, <rotation_matrix>

For some demo scenes, we include a view file named demo_video.csv which is recorded from real viewing by user.

${MODEL_FOLDER} should contains two model files: fovea.tar for the foveal model and periph.tar for the periph model. We also contain pretrained models for several demo scenes in pretrained/ for test.

-f ${FPS} argument specifies the FPS of output video. Without this argument, the script will generate a sequence of images instead.

The outputs are two videos (or two folders of images) stored in the same folder as ${VIEW_FILE}. The video file (or folder of images) whose name contains hint is the output of views with overlay hints to indicate where the gazes are.

We implemented a simple real-time interactive rendering program based on OpenGL, CUDA and TensorRT. To play with that, first follow the guide to install the dependencies and compile the program, then export pretrained models. After that, run the program by:

cd cpp/bin && ./view_gl ../nets/${SCENE}

The optional command-line arguments are:

-h: print the help information and exit

-s|-fs: render stereo view with/without stereo adaption (Section 3.2.2)

-p: print the performance information of every frame to the standard output

With the rendering window focused, you can interactively change the view and gaze by keyboard and mouse:

Gaze positions:

For stereo view, there are two gazes in the window (for the left view and right view). Click the mouse in left view region will move the left gaze directly to the clicked position. The right gaze will be move correspondingly according to current disparity value and vice versa.

Disparity:

For stereo view with stereo adaption, misalignment will be obvious when the left and right gazes are not focused on the same object (i.e. wrong disparity). You can control the disparity by scrolling the wheel to see the changing of this effect as well as the periphery shifting as described in Section 3.2.2. You can also use the key "S" to intuitively compare between the effects of rendering with/without periphery shifting.

The program can only load TensorRT workspace file. So we include a tool script to convert the PyTorch model to ONNX model as well as TensorRT workspace:

python export.py ${MODEL_FILE} --batch-size ${BATCH_SIZE} --trt [--outdir ${OUT_DIR}]

The tool first export the PyTorch model to an ONNX model file and an INI file containing the configurations required by view_gl program. With --trt argument specified, it further load the ONNX model and build a TensorRT workspace using trtexec.

The ${BATCH_SIZE} should be 51429 for foveal model and 108828 for periphery model.

Outputs will be stored in ${OUT_DIR}/ laying besides the PyTorch model file. The default output folder is exported.

For example, to export the pretrained model for barbershop, use the following command:

export SCENE=barbershop
python export.py pretrained/${SCENE}/fovea.tar --batch-size 51429 --trt && \
python export.py pretrained/${SCENE}/periph.tar --batch-size 108828 --trt

After run, pretrained/barbershop/ will contain following files:

pretrained/barbershop
|-- fovea.tar
|-- periph.tar
|-- exported       # The output folder
    |-- fovea.ini
    |-- fovea.onnx
    |-- fovea.trt
    |-- periph.ini
    |-- periph.onnx
    |-- periph.trt

After exported the foveal and periphery models, move all .trt and .ini files in the output folder to a new folder under cpp/nets. Make sure the file names are fovea.trt, fovea.ini, periph.trt and periph.ini:

mkdir cpp/nets/${SCENE} && \
cd pretrained/${SCENE}/exported && \
mv *.ini *.trt ../../../cpp/nets/${SCENE}/ && \
cd ../../..

No. Currently we only support training on a single GPU. If you have multiple GPUs installed in your computer, the training will default run on the first one. You can specify which GPU to use by:

export CUDA_VISIBLE_DEVICES=X

before runing the scripts.

You can use the following command to start your training in background:

nohup python train.py ${ARGUMENTS} > /dev/null 2>&1 &

This command will return immediately and disable the output of messages to the terminal. You can check the log file for the training progress.

Please cite as

@article{deng2022fovnerf,
  title = {FoV-NeRF: Foveated Neural Radiance Fields for Virtual Reality},
  author = {Deng, Nianchen and He, Zhenyi and Ye, Jiannan and Duinkharjav, Budmonde and Chakravarthula, Praneeth and Yang, Xubo and Sun, Qi},
  year = 2022,
  journal = {IEEE Transactions on Visualization and Computer Graphics},
  volume = {},
  number = {},
  pages = {1--11},
  doi = {10.1109/TVCG.2022.3203102}
}