Official implementation of Instruments as Queries for Audio-Visual Sound Separation.

CVPR Paper Project Page Video

This paper has been accepted by CVPR 2023.

We re-formulate visual-sound separation task and propose Instrument as Query (iQuery) with a flexible query expansion mechanism. Our approach ensures cross-modal consistency and cross-instrument disentanglement. We utilize "visually named" queries to initiate the learning of audio queries and use cross-modal attention to remove potential sound source interference at the estimated waveforms. To generalize to a new instrument or event class, drawing inspiration from the text-prompt design, we insert an additional query as an audio prompt while freezing the attention mechanism.

Create a conda environment and install dependencies:

git clone https://github.com/JiabenChen/iQuery.git
cd iQuery

conda create --name iQuery python=3.8
conda activate iQuery

# Install the according versions of torch and torchvision
pip install torch==1.12.0+cu102 torchvision==0.13.0+cu102 torchaudio==0.12.0 --extra-index-url 

# Install detectron2
git clone [email protected]:facebookresearch/detectron2.git
cd detectron2
pip install -e .

# Install required packages
pip install -U opencv-python
pip install -r requirements.txt

1. Download datasets.

   a. Download MUSIC and MUSIC-21 dataset from: https://github.com/roudimit/MUSIC_dataset (The original MUSIC datasets provides .json files, you can refer to iQuery/scripts/download_videos.py to download videos)

   b. Download AVE dataset from: https://github.com/YapengTian/AVE-ECCV18

2. Preprocess videos. You can do it in your own way as long as the index files are similar.

   a. Extract frames and waveforms (11025Hz for MUSIC/MUSIC-21, and 22000Hz for AVE). (You can refer to iQuery/scripts/extract_audio.py and iQuery/scripts/extract_frames.py)

   b. Detect objects in video frames. For MUSIC dataset, we used object detector trained by Ruohan used in his Cosep project (see CoSep repo). For MUSIC-21 and AVE, we used a pre-trained Detic detector (see Detic repo) to detect the 10 more instruments in MUSIC-21 dataset and 28 event classes in AVE dataset. The detected objects for each video are stored in a .npy file.

   c. Extract motion features. We adopt the pretrained I3D video encoder used in FAME repo. The extracted motion features are stored in a .npy file. (You could refer to iQuery/scripts/extract_motion.py)

3. Data splits. We created index files as .csv for training and testing. Index files for MUSIC/MUSIC-21/AVE datasets can be found at iQuery/data/MUSIC, iQuery/data/MUSIC_21 and iQuery/data/AVE, respectively.

4. Directory structure. We have following directory structure for data:

   data
   ├── audio
   |   ├── acoustic_guitar
   │   |   ├── M3dekVSwNjY.wav
   │   |   ├── ...
   │   ├── trumpet
   │   |   ├── STKXyBGSGyE.wav
   │   |   ├── ...
   │   ├── ...
   |
   └── frames
   |   ├── acoustic_guitar
   │   |   ├── M3dekVSwNjY.mp4
   │   |   |   ├── 000001.jpg
   │   |   |   ├── ...
   │   |   ├── ...
   │   ├── trumpet
   │   |   ├── STKXyBGSGyE.mp4
   │   |   |   ├── 000001.jpg
   │   |   |   ├── ...
   │   |   ├── ...
   │   ├── ...
   |
   └── detection_results
   |   ├── acoustic_guitar
   │   |   ├── M3dekVSwNjY.mp4.npy
   │   |   ├── ...
   │   ├── trumpet
   │   |   ├── STKXyBGSGyE.mp4.npy
   │   |   ├── ...
   │   ├── ...
   |
   └── motion_features
   |   ├── acoustic_guitar
   │   |   ├── M3dekVSwNjY.mp4.npy
   │   |   ├── ...
   │   ├── trumpet
   │   |   ├── STKXyBGSGyE.mp4.npy
   │   |   ├── ...
   │   ├── ...
   

1. Train the full iQuery model a. on MUSIC dataset

   cd code
   bash scripts/train_music.sh

   b. on MUSIC-21 dataset

   cd code
   bash scripts/train_music21.sh

   c. on AVE dataset

   cd code
   bash scripts/train_ave.sh

2. Train iQuery model without motion

   cd code
   bash scripts/train_nomotion.sh

Please use the following script to evaluate iQuery's performance, you should only modify the pretrained model id and main file name in the script:

cd code
bash scripts/evaluate.sh

Pretrained checkpoints of iQuery can be found at: https://drive.google.com/file/d/17jzOwsTDBaHVw18eaErNZtvGYaB7TN20/view?usp=sharing

@inproceedings{chen2023iquery,
  title={iQuery: Instruments as Queries for Audio-Visual Sound Separation},
  author={Chen, Jiaben and Zhang, Renrui and Lian, Dongze and Yang, Jiaqi and Zeng, Ziyao and Shi, Jianbo},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={14675--14686},
  year={2023}
}

Our code is based on the implementation of Sound-of-Pixels, CCoL, and AVE. We sincerely thanks those authors for their great works. If you use our codes, please also consider cite their nice works.

If you have any question about this project, please feel free to contact [email protected].