Dataset Card for VGGSound-1000

This is a FiftyOne dataset with 1000 samples.

This FiftyOne dataset contains a random sample of 1,000 clips (take(1000, seed=51)) from the first shard (vggsound_00.tar.gz, 10,000 clips) of the 199,467-clip VGGSound release, not the full dataset. The full 10,000-clip shard is at harpreetsahota/VGGSound.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("harpreetsahota/VGGSound-1000")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

VGG-Sound is an audio-visual dataset of 10-second clips extracted from YouTube videos, each labeled with one of 309 sound classes. The sound source in each clip is visually evident (audio-visual correspondence), and labels were generated by an automated pipeline (visual verification, audio verification and iterative noise filtering) rather than by human annotation. The full dataset has ~200k clips and 309 classes according to the paper; this FiftyOne version holds a random sample of 1,000 of the 10,000 clips of the first archive shard. It also carries EmbeddingGemma 2 embeddings of the frames, the audio and both together, with similarity and visualization runs and uniqueness and representativeness scores computed from them (see Dataset Structure).

• Curated by: Honglie Chen, Weidi Xie, Andrea Vedaldi, Andrew Zisserman (Visual Geometry Group, University of Oxford)
• Funded by: [More Information Needed]
• Shared by: [More Information Needed]
• Language(s): en
• License: CC BY 4.0 (the copyright of the videos remains with the original owners)

Dataset Sources

• Repository: https://huggingface.co/datasets/Loie/VGGSound (original release: https://www.robots.ox.ac.uk/~vgg/data/vggsound/)
• Paper: https://arxiv.org/abs/2004.14368
• Demo: [More Information Needed]

Uses

Direct Use

Audio recognition and audio-visual learning, e.g. audio classification and multi-modal audio-visual analysis, where the sound source is visible in the video. The paper also suggests tasks such as audio grounding.

Out-of-Scope Use

[More Information Needed]

Dataset Structure

This is a flat FiftyOne video dataset (media_type="video") with 1,000 samples: one sample per 10-second .mp4 clip (h264 video with an aac audio track at 44.1 kHz). Clips have variable resolution (about 200x360 up to 1280x720) and frame rate. Each sample is also tagged train (930 samples) or test (70 samples) following the split column of the source CSV.

Field	FiftyOne type	Description
filepath	StringField	Path to the .mp4 clip
tags	ListField(StringField)	train or test, copied from the source CSV split
metadata	VideoMetadata	Duration, frame rate, resolution, etc. computed with compute_metadata()
youtube_id	StringField	YouTube video ID from the source CSV (verbatim)
start_sec	IntField	Start time in seconds of the 10-second clip within the YouTube video (verbatim)
split	StringField	train or test (verbatim from the source CSV)
ground_truth	Classification	Sound class of the clip (266 distinct labels in this sample)
eg2_video	VectorField	768-d EmbeddingGemma 2 vector of the video frames only (the sound is ignored). Derived, not from the source
eg2_audio	VectorField	768-d EmbeddingGemma 2 vector of the soundtrack (first 30 seconds). Derived
eg2_av	VectorField	768-d EmbeddingGemma 2 vector of frames and soundtrack together, as one vector. Derived
eg2_video_uniqueness, eg2_audio_uniqueness, eg2_av_uniqueness	FloatField	Uniqueness score of the clip among the 1,000, computed from the matching vector field. Derived
eg2_video_representativeness, eg2_audio_representativeness, eg2_av_representativeness	FloatField	Representativeness score of the clip, computed from the matching vector field. Derived

Embeddings: the three vector fields were computed with EmbeddingGemma 2, which puts text, images, video and audio in one 768-dimensional space, so a text query can be compared with any of them. Video frames were sampled at 1 frame per second, with at most 32 frames per clip, spread over the clip. Audio is truncated to the first 30 seconds. All 1,000 clips have a vector in each field, and every vector is unit length.

Brain runs: the Hub repository carries 6 FiftyOne brain runs, 2 per vector field (eg2_video, eg2_audio, eg2_av):

Brain key	Method	Notes
eg2_video, eg2_audio, eg2_av	similarity, lancedb backend	Cosine metric. The model google/embeddinggemma-2 and its settings are recorded on each run, so text search works
eg2_video_viz, eg2_audio_viz, eg2_av_viz	visualization, umap	2 dimensions, 15 neighbors, min_dist 0.1, euclidean metric, seed 51

The uniqueness and representativeness scores (cluster-center method for representativeness) are stored in the eg2_*_uniqueness and eg2_*_representativeness sample fields listed above. Their brain runs were not exported to the Hub repository, so only the scores are available.

The LanceDB tables behind the three similarity runs are not part of the Hub repository: they live outside the dataset. The vectors are in the eg2_* fields, and the tables can be rebuilt from them.

Label types: each clip has exactly one flat sound label (the paper's label set has no hierarchy), so it is stored as a single fo.Classification rather than Classifications.

dataset.info: source repo URL, paper URL, license, a note that this is a 1,000-clip sample of the vggsound_00 shard only, and csv_rows_total (199,467), the number of rows in the full source CSV.

Parsing decisions:

• Labels, splits and start times come from vggsound.csv (headerless: youtube_id, start_sec, label, split). Clip filenames are {youtube_id}_{start_sec:06d}.mp4, and each CSV row was matched to a clip by that name. All 10,000 clips in the shard matched a CSV row and none were orphaned.
• Only rows whose clip is present on disk were imported.
• The clips were moved out of the archive's nested path into a flat videos/ directory.
• Some labels contain commas (for example female speech, woman speaking), so the CSV is parsed with a CSV reader rather than by splitting on commas.
• The 1,000 clips are a random sample (take(1000, seed=51)) of the 10,000 clips of the vggsound_00 shard.
• The eg2_* vector fields were computed on all 10,000 clips of the shard with no failures, and the 1,000 sampled clips keep their vectors. The brain runs and the uniqueness and representativeness scores were recomputed on the 1,000 clips from those vectors, so they describe this sample only.

Dataset Creation

Curation Rationale

The authors wanted a large-scale audio dataset of sounds "in the wild" with low label noise and guaranteed audio-visual correspondence (the sound source is visible), collected without the extensive human effort used to build earlier manually curated datasets such as AudioSet.

Source Data

Data Collection and Processing

Videos were obtained from YouTube through a multi-stage pipeline: (1) a list of sound classes and candidate videos was built (about 600 classes, 1M videos); (2) visual verification with image classification kept videos where the sound source is visible (470 classes, 550k videos); (3) audio verification filtered out ambient noise (390 classes, 260k videos); (4) iterative noise filtering produced the final set (309 classes, 200k videos). Classes with fewer than 100 videos were removed, and there are no more than 2 clips per video. Each clip is 10 seconds long.

Who are the source data producers?

The creators and uploaders of the YouTube videos from which the clips were extracted. [More Information Needed] on any further demographic details.

Annotations

Annotation process

Labels are generated by the automated pipeline described above, which the authors say avoids the need for human annotation; manual input is needed only at a few well-defined points.

Who are the annotators?

Annotations are produced automatically by the pipeline, not by human annotators.

Personal and Sensitive Information

[More Information Needed]

Citation

BibTeX:

@InProceedings{Chen20,
  author       = "Honglie Chen and Weidi Xie and Andrea Vedaldi and Andrew Zisserman",
  title        = "VGGSound: A Large-scale Audio-Visual Dataset",
  booktitle    = "International Conference on Acoustics, Speech and Signal Processing (ICASSP)",
  year         = "2020",
}

APA:

Chen, H., Xie, W., Vedaldi, A., & Zisserman, A. (2020). VGGSound: A large-scale audio-visual dataset. In International Conference on Acoustics, Speech and Signal Processing (ICASSP).

More Information

The original dataset lists a train/test split in vggsound.csv; the same split is preserved here as sample tags and the split field. The full 10,000-clip shard is at harpreetsahota/VGGSound.

Dataset Card Authors

[More Information Needed]

Dataset Card Contact

[More Information Needed]