[CVPR-2023] Audio-Visual Grouping Network for Sound Localization from Mixtures

AVGN is a new approach for disentangling category-wise semantic features for each source from the mixture and image to localize multiple sounding sources simultaneously.

Environment

To setup the environment, please simply run

pip install -r requirements.txt

Datasets

MUSIC

Data can be downloaded from Sound of Pixels

VGG-Instruments

Data can be downloaded from Mix and Localize: Localizing Sound Sources in Mixtures

VGG-Sound Source

Data can be downloaded from Localizing Visual Sounds the Hard Way

Train

For training the AVGN model, please run

python train.py --multiprocessing_distributed \
    --train_data_path /path/to/vgginstruments/train/ \
    --test_data_path /path/to/vgginstruments/ \
    --test_gt_path /path/to/vgginstruments/anno/ \
    --experiment_name vgginstruments_multi_avgn \
    --model 'avgn' \
    --trainset 'vgginstruments_multi' --num_class 37 \
    --testset 'vgginstruments_multi' \
    --epochs 100 \
    --batch_size 128 \
    --init_lr 0.0001 \
    --attn_assign soft \
    --dim 512 \
    --depth_aud 3 \
    --depth_vis 3

Test

For testing and visualization, simply run

python test.py --test_data_path /path/to/vgginstruments/ \
    --test_gt_path /path/to/vgginstruments/anno/ \
    --model_dir checkpoints \
    --experiment_name vgginstruments_multi_avgn \
    --model 'avgn' \
    --testset 'vgginstruments_multi' \
    --alpha 0.3 \
    --attn_assign soft \
    --dim 512 \
    --depth_aud 3 \
    --depth_vis 3

Citation

If you find this repository useful, please cite our paper:

@inproceedings{mo2023audiovisual,
  title={Audio-Visual Grouping Network for Sound Localization from Mixtures},
  author={Mo, Shentong and Tian, Yapeng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}