Code for Acoustic Field Reconstruction

[NeurIPS'24 Spotlight] AVR: Acoustic Volume Rendering for Neural Impulse Response Fields

[NeurIPS'25 Paper] Versa: Resounding Acoustic Fields with Reciprocity

[AVR arXiv] [AVR Website] [AcoustiX Code] [BibTex]

[Versa arXiv] [Versa Website]

Contents

This repo contains the official implementation for AVR. For our simulator: AcoustiX Code, please use another repo.

• Abstract
• Codebase
• Usage
• Citation

Updates

• 2025.12.18: Add Versa repo.

Codebase

Below is the instructions on how to install and set up the project.

Requirements (in addition to the common python stack)

• pytorch
• numpy
• scipy
• matplotlib
• librosa
• auraloss

In addition to above common python packages, we also use tinycudann to speed up the ray sampling. Installation of this repo and python extension is shown below.

# install tiny-cuda-nn PyTorch extension
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

Project structure

AVR/
├── config_files/          # training and testing config files
│   ├── avr_raf_furnished.yml
│   ├── avr_raf_empty.yml
│   ├── avr_meshrir.yml
│   └── avr_simu.yml       
├── logs/                  # Log files
│   ├── meshrir            # Meshrir logs
│   ├── RAF                # RAF logs
│   └── simu               # Simulation logs
├── tensorboard_logs/      # TensorBoard log files
├── data/                  # Dataset 
├── utils/                 # Utility scripts
│   ├── criterion.py       # Loss functions   
│   ├── logger.py          
│   ├── spatialization.py  # Audio spatialization        
│   └── metric.py          # Metrics calculation.
├── tools/                 # Tools to create datasets and more   
│   └── meshrir_split.py   # Create meshrir dataset split
├── avr_runner.py          # AVR runner
├── datasets_loader.py     # dataloader for different datasets
├── model.py               # network
├── renderer.py            # acoustic rendering file
├── README.md              # Project documentation
└── .gitignore             # Git ignore file

Example Usage

• Train AVR on RAF-Furnished dataset

python avr_runner.py --config ./config_files/avr_raf_furnished.yml --dataset_dir ./data/RAF/FurnishedRoomSplit

• Train AVR on RAF-Empty dataset

python avr_runner.py --config ./config_files/avr_raf_empty.yml --dataset_dir ./data/RAF/EmptyRoomSplit

• Train AVR on MeshRIR dataset

python avr_runner.py --config ./config_files/avr_meshrir.yml  --dataset_dir ./data/MeshRIR

Create your own dataset

MeshRIR dataset: Refer to Create Meshrir Dataset Instructions

Results and Visualizations

We show some visualization results from our paper:

We show the impusle response frequency spatial distribution with ground truth and different baseline methods. This is a bird-eye view of signal distributions.

We show the estimated impulse response from different datasets and methods.

Citation

If you find this project to be useful for your research, please consider citing the paper.

@inproceedings{lanresounding,
  title={Resounding Acoustic Fields with Reciprocity},
  author={Lan, Zitong and Hao, Yiduo and Zhao, Mingmin},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}
}

@inproceedings{lanacoustic,
  title={Acoustic Volume Rendering for Neural Impulse Response Fields},
  author={Lan, Zitong and Zheng, Chenhao and Zheng, Zhiwei and Zhao, Mingmin},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}
}