Long YouTube Videos With Dense Captioning For Scene Transitions

87.8K long, untrimmed YouTube videos averaging 486 seconds each provide 11.8K hours of footage. The dataset contains 1.3M segmented video clips, each with dense and detailed captions, created by IVLLab and updated in December 2024.

Use Cases

• Train video captioning models on the dense, detailed descriptions provided for each of the 1.3M segmented video clips.
• Analyze scene transition frequency and patterns across the 87.8K long, untrimmed videos.
• Develop scene recognition models using the thorough captions that describe the entire scene context for each segment.
• Benchmark models for understanding long-form video content using the 11.8K hours of video duration.

Strengths

• Large scale of 1.3M segmented video clips for training.
• Long video context with 87.8K source videos averaging 486 seconds each.
• High-quality, dense and detailed captions for every video segment.

Limitations

• Sourced solely from YouTube, which may introduce platform-specific biases in content and style.
• The raw description notes frequent scene transitions, which could indicate challenging, discontinuous content for some models.
• Dataset size, file formats, and specific column structure are unknown from the provided input.

Provenance

Source

YouTube videos.

Collection Method

Sourced from long and untrimmed YouTube videos, with dense captioning applied.

Time Range

null

Freshness

Last updated in December 2024.

Geography

null