lemonweed6312/VGGSound-Duet-Mask

VGGSound-Duet^Mask

VGGSound-Duet^Mask provides pixel-level segmentation masks for evaluating dual-source audio-visual localisation. It accompanies SCAV, introduced in Whence the Voice? Self-supervised Dual-source Audio-Visual Localisation via Selective Convergence (ECCV 2026).

For more information, visit the project page, read the paper, or see the official code.

Overview

Conventional sound-source localisation benchmarks frequently use bounding boxes, which include background pixels and can overestimate localisation quality. VGGSound-Duet^Mask replaces box-level evaluation with foreground segmentation masks so that predicted heatmaps can be compared with the visible sound-producing regions more precisely.

Key features

• Pixel-level masks for the VGGSound-Duet benchmark.
• Designed for mask-based evaluation of dual-source audio-visual localisation.
• Supports the 3,951 currently accessible VGGSound-Duet test pairs spanning 220 sound categories described in the SCAV paper.
• Contains 4,512 individual source-frame mask files in the current Hugging Face repository. Individual masks are reused when constructing the dual-source test pairs.
• Masks were generated with the Segment Anything Model (SAM) and stored as NumPy arrays serialized with Python pickle.

This repository contains the mask annotations only. It does not redistribute the source videos, audio, extracted frames, optical flow, VGGSound metadata, or SCAV checkpoints.

Download

The repository is public and does not require an access token.

Hugging Face CLI

pip install -U "huggingface_hub[cli]"
hf download lemonweed6312/VGGSound-Duet-Mask \
  --repo-type dataset \
  --local-dir ./VGGSound-Duet-Mask

Python

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="lemonweed6312/VGGSound-Duet-Mask",
    repo_type="dataset",
    local_dir="./VGGSound-Duet-Mask",
)

To download one mask only:

from huggingface_hub import hf_hub_download

mask_path = hf_hub_download(
    repo_id="lemonweed6312/VGGSound-Duet-Mask",
    repo_type="dataset",
    filename="masks/-0BIyqJj9ZU_000030.pkl",
)

Dataset structure

VGGSound-Duet-Mask/
├── README.md
└── masks/
    ├── -0BIyqJj9ZU_000030.pkl
    ├── -2-wdcN5vOw_000017.pkl
    └── ...

Each filename follows this pattern:

<youtube_video_id>_<frame_index>.pkl

The frame index is zero-padded to six digits. Each pickle contains one NumPy mask array. The evaluation code accepts either a two-dimensional array (H, W) or a single-channel array (H, W, 1). Spatial dimensions follow the corresponding source frame and may vary. Masks are interpreted as binary using a threshold of 0.5.

For example, -0BIyqJj9ZU_000030.pkl contains a float64 NumPy array of shape (240, 320, 1) with values 0 and 1. Its size of approximately 615 kB is expected: an uncompressed 240 × 320 × 1 float64 array alone requires 614,400 bytes.

Reading a mask

import pickle
from pathlib import Path

import numpy as np

mask_path = Path("VGGSound-Duet-Mask/masks/-0BIyqJj9ZU_000030.pkl")

with mask_path.open("rb") as file:
    mask = pickle.load(file)

if mask.ndim == 3:
    mask = mask[:, :, 0]

mask = (mask > 0.5).astype(np.float32)
print(mask.shape, mask.dtype, mask.min(), mask.max())

The pretrained weights are available in the SCAV model repository. See the official code README for the remaining data preparation and evaluation requirements.

Dataset details

• Developed by: Han Hu, Dongheng Lin, Yuqi Hou, Haotian Li, Hyung Jin Chang, and Jianbo Jiao
• Institution: The MIx Group, University of Birmingham
• Associated paper: Whence the Voice? Self-supervised Dual-source Audio-Visual Localisation via Selective Convergence
• Conference: ECCV 2026
• Task: dual-source audio-visual localisation evaluation
• Annotation type: foreground segmentation mask
• Number of mask files in this repository: 4,512
• Benchmark coverage reported in the paper: 3,951 currently accessible dual-source test pairs across 220 sound categories

Limitations and responsible use

• The masks are automatically generated with SAM and may contain segmentation errors, omissions, or ambiguous object boundaries.
• Dataset coverage depends on the availability of the original VGGSound videos; some source videos are no longer available on YouTube.
• The annotations identify visible foreground regions used for benchmark evaluation. They should not be interpreted as exhaustive semantic segmentation labels.
• Users are responsible for complying with the terms and licenses of VGGSound, YouTube, and any source media they obtain separately.

Citation

If you use VGGSound-Duet^Mask or the SCAV checkpoints, please cite:

@inproceedings{hu2026scav,
  title     = {Whence the Voice? Self-supervised Dual-source Audio-Visual Localisation via Selective Convergence},
  author    = {Hu, Han and Lin, Dongheng and Hou, Yuqi and Li, Haotian and Chang, Hyung Jin and Jiao, Jianbo},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}