<div align="center">

<h2>AV-NeRF: Learning Neural Fields for Real-World Audio-Visual Scene Synthesis</h2>

_**[Susan Liang](https://liangsusan-git.github.io/), [Chao Huang](https://wikichao.github.io/), [Yapeng Tian](https://www.yapengtian.com/), [Anurag Kumar](https://anuragkr90.github.io/), [Chenliang Xu](https://www.cs.rochester.edu/~cxu22/)**_

</div>

### RWAVS Dataset
We provide the Real-World Audio-Visual Scene (RWAVS) Dataset.

1. The dataset can be downloaded from this Hugging Face repository.

2. After you download the dataset, you can decompress the `RWAVS_Release.zip`.
    ```
    unzip RWAVS_Release.zip
    cd release/
    ```

3. The data is organized with the following directory structure.
    ```
    ./release/
    ├── 1
    │   ├── binaural_syn_re.wav
    │   ├── feats_train.pkl
    │   ├── feats_val.pkl
    │   ├── frames
    │   │   ├── 00001.png
    |   |   ├── ...
    │   │   ├── 00616.png
    │   ├── source_syn_re.wav
    │   ├── transforms_scale_train.json
    │   ├── transforms_scale_val.json
    │   ├── transforms_train.json
    │   └── transforms_val.json
    ├── ...
    ├── 13
    └── position.json
    ```

    The dataset contains 13 scenes indexed from 1 to 13. For each scene, we provide
    * `transforms_train.json`: camera poses for training.
    * `transforms_val.json`: camera poses for evaluation. We split the data into `train` and `val` subsets with 80% data for training and the rest for evaluation.
    * `transforms_scale_train.json`: normalized camera poses for training. We scale 3D coordindates to $[-1, 1]^3$.
    * `transforms_scale_val.json`: normalized camera poses for evaluation.
    * `frames`: corresponding video frames for each camera pose.
    * `source_syn_re.wav`: single-channel audio emitted by the sound source.
    * `binaural_syn_re.wav`: two-channel audio captured by the binaural microphone. We synchronize `source_syn_re.wav` and `binaural_syn_re.wav` and resample them to $22050$ Hz.
    * `feats_train.pkl`: extracted vision and depth features at each camera pose for training. We rely on V-NeRF to synthesize vision and depth images for each camera pose. We then use a pre-trained encoder to extract features from rendered images.
    * `feats_val.pkl`: extracted vision and depth features at each camera pose for inference.
    * `position.json`: normalized 3D coordinates of the sound source.

    Please note that some frames may not have corresponding camera poses because COLMAP fails to estimate the camera parameters of these frames.

### Citation
```bib
@inproceedings{liang23avnerf,
 author = {Liang, Susan and Huang, Chao and Tian, Yapeng and Kumar, Anurag and Xu, Chenliang},
 booktitle = {Conference on Neural Information Processing Systems (NeurIPS)},
 title = {AV-NeRF: Learning Neural Fields for Real-World Audio-Visual Scene Synthesis},
 year = {2023}
}
```

### Contact
If you have any comments or questions, feel free to contact [Susan Liang](mailto:sliang22@ur.rochester.edu) and [Chao Huang](mailto:chuang65@ur.rochester.edu).