Sound Source Localization is All About Alignment (ICCV’23)

Official PyTorch implementation of our following papers:

Sound Source Localization is All About Cross-Modal Alignment

Arda Senocak*, Hyeonggon Ryu*, Junsik Kim*, Tae-Hyun Oh, Hanspeter Pfister, Joon Son Chung (* Equal Contribution)

ICCV 2023

Toward Interactive Sound Source Localization: Better Align Sight and Sound!

Arda Senocak*, Hyeonggon Ryu*, Junsik Kim*, Tae-Hyun Oh, Hanspeter Pfister, Joon Son Chung (* Equal Contribution)

TPAMI 2025

Index

• Overview
• Interactive Synthetic Sound Source (IS3) Dataset
• Environment
• Model Checkpoints
• Inference
• Training
• Citation

Overview

Interactive Synthetic Sound Source (IS3) Dataset

IS3 dataset is available here

The IS3 data is organized as follows:

Note that in IS3 dataset, each annotation is saved as a separate file. For example; the sample accordion_baby_10467 image contains two annotations for accordion and baby objects. These annotations are saved as accordion_baby_10467_accordion and accordion_baby_10467_baby for straightforward use. You can always project bounding boxes or segmentation maps onto the original image to see them all at once.

images and audio_waw folders contain all the image and audio files respectively.

IS3_annotation.json file contains ground truth bounding box and category information of each annotation.

gt_segmentation folder contains segmentation maps in binary image format for each annotation. You can query the file name in IS3_annotation.json to get semantic category of each segmentation map.

Environment

Model Checkpoints

The model checkpoints are available for the following experiments:

Training Set	Test Set	Model Type	Performance (cIoU)	Checkpoint
VGGSound-144K	VGG-SS	NN w/ Sup. Pre. Enc.	39.94	Link
VGGSound-144K	VGG-SS	NN w/ Self-Sup. Pre. Enc.	39.16	Link
VGGSound-144K	VGG-SS	NN w/ Sup. Pre. Enc. Pre-trained Vision	41.42	Link
Flickr-SoundNet-144K	Flickr-SoundNet	NN w/ Sup. Pre. Enc.	85.20	Link
Flickr-SoundNet-144K	Flickr-SoundNet	NN w/ Self-Sup. Pre. Enc.	84.80	Link
Flickr-SoundNet-144K	Flickr-SoundNet	NN w/ Sup. Pre. Enc. Pre-trained Vision	86.00	Link

Inference

Put checkpoint files into the 'checkpoints' directory:

inference
│
└───checkpoints
│       ours_sup_previs.pth.tar
│       ours_sup.pth.tar
│       ours_selfsup.pth.tar
│   test.py
│   datasets.py
│   model.py

To evaluate a trained model run

python test.py --testset {testset_name} --pth_name {pth_name}

Test Set	testset_name
VGG-SS	vggss
Flickr-SoundNet	flickr
IS3	is3

Evaluate other methods

Simply save the checkpoint files from the methods as '{method_name}_{put_your_own_message}.pth', such as 'ezvsl_flickr.pth'. We have already handled the trivial settings.

Paper title	pth_name must contains
Localizing Visual Sounds the Hard Way (CVPR 21) [Paper]	lvs
Localizing Visual Sounds the Easy Way (ECCV 22) [Paper]	ezvsl
A Closer Look at Weakly-Supervised Audio-Visual Source Localization (NeurIPS 22) [Paper]	slavc
Exploiting Transformation Invariance and Equivariance for Self-supervised Sound Localisation (ACMMM 22) [Paper]	ssltie
Learning Audio-Visual Source Localization via False Negative Aware Contrastive Learning (CVPR 23) [Paper]	fnac

Example

python test.py --testset flickr --pth_name ezvsl_flickr.pth

Training

Training code is coming soon!

Citation

If you find this code useful, please consider giving a star ⭐ and citing us:

@inproceedings{senocak2023sound,
  title={Sound source localization is all about cross-modal alignment},
  author={Senocak, Arda and Ryu, Hyeonggon and Kim, Junsik and Oh, Tae-Hyun and Pfister, Hanspeter and Chung, Joon Son},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={7777--7787},
  year={2023}
}

If you use this dataset, please consider giving a star ⭐ and citing us:

@article{senocak2025align,
  author={Senocak, Arda and Ryu, Hyeonggon and Kim, Junsik and Oh, Tae-Hyun and Pfister, Hanspeter and Chung, Joon Son},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Toward Interactive Sound Source Localization: Better Align Sight and Sound!}, 
  year={2025},
}