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.
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
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
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
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
- 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.
NeRF2 is MIT-licensed. The license applies to the pre-trained models and datasets as well.
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}
}
Some code snippets are borrowed from nerf-pytorch and NeuS.