VGGSound: Audio-Visual Correspondences from YouTube Videos

VGGSound is an audio-visual dataset consisting of short clips extracted from videos uploaded to YouTube. It contains over 200,000 videos spanning more than 310 acoustic classes. The dataset was created by researchers at the University of Oxford and was last updated on the platform in March 2023.

Use Cases

• Train audio classification models based on the 310+ acoustic classes.
• Develop multimodal learning models based on the audio-visual correspondences.
• Research acoustic environments and noise characteristics based on the described real-world applications.

Strengths

• Contains over 200,000 video clips.
• Spans more than 310 distinct acoustic classes.
• Clips are sourced from real-world YouTube videos, suggesting diverse acoustic environments.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2023-03-26 13:25:40; freshness should be verified.

Provenance

Source

University of Oxford VGG group.

Collection Method

Clips extracted from videos uploaded to YouTube.

Freshness

Last updated 2023-03-26 13:25:40.