NeRF2: Neural Radio-Frequency Radiance Fields

Thank you for your interest in our work. This repository maintains code for NeRF², recognized as the Best Paper Runner-Up at ACM MobiCom 2023. NeRF² is a physical-layer neural network capable of accurately predicting signal characteristics at any location based on the position of a transmitter. By integrating learned statistical models with physical ray tracing, NeRF² creates synthetic datasets ideal for training application-layer neural networks. This technology also demonstrates potential in indoor localization and 5G MIMO channel prediction, showcasing an fusion of wireless communication and AI.

Project | Paper | Datasets

RFID spectrum / BLE / MIMO prediction

Datasets and pretrained models are available at Here.

The datasets are organized as follows:

NeRF2-Dataset
|-- BLE   # BLE RSSI Prediction Dataset
    |-- rssi-ckpts-1.tar         # pretrained model
    |-- rssi-dataset-1.tar.gz    # rssi dataset
|-- MIMO   # MIMO CSI Prediction Dataset
    |-- csi-ckpts-1.tar          # pretrained model
    |-- csi-dataset-1.tar.gz     # csi dataset
|-- RFID   # RFID Spectrum Prediction Dataset
    |-- s23-ckpts.tar            # pretrained model
    |-- s23-dataset.tar.gz       # spectrum dataset

Running

Spectrum prediction

training the model

python nerf2_runner.py --mode train --config configs/rfid-spectrum.yml --dataset_type rfid --gpu 0

Inference the model

python nerf2_runner.py --mode test --config configs/rfid-spectrum.yml --dataset_type rfid --gpu 0

RSSI prediction

training the model

python nerf2_runner.py --mode train --config configs/ble-rssi.yml --dataset_type ble --gpu 0

Inference the model

python nerf2_runner.py --mode test --config configs/ble-rssi.yml --dataset_type ble --gpu 0

MRI

python baseline/mri.py

CSI prediction

training the model

python nerf2_runner.py --mode train --config configs/mimo-csi.yml --dataset_type mimo --gpu 0

Inference the model

python nerf2_runner.py --mode test --config configs/mimo-csi.yml --dataset_type mimo --gpu 0

To-Do List

• CGAN RSSI prediction baseline
• Release more datasets
• Instruction of preparing own datasets
• Implementation on Taichi to speed up the code

Please stay tuned for updates and feel free to reach out if you have any questions or need further information.

License

NeRF² is MIT-licensed. The license applies to the pre-trained models and datasets as well.

Citation

If you find the repository is helpful to your project, please cite as follows:

@inproceedings{zhao2023nerf2,
    author = {Zhao, Xiaopeng and An, Zhenlin and Pan, Qingrui and Yang, Lei},
    title = {NeRF2: Neural Radio-Frequency Radiance Fields},
    booktitle = {Proc. of ACM MobiCom '23},
    pages = {1--15},
    year = {2023}
}

Acknowledgment

Some code snippets are borrowed from nerf-pytorch and NeuS.