HuggingFace Releases High-Quality Video Dataset FineVideo

To address the data shortage that has been a bottleneck for open-source video AI development, HuggingFace has launched the FineVideo dataset. This dataset consists of a total of 43,000 videos (approximately 3,400 hours), and includes not just raw footage but rich annotations such as detailed descriptions, narrative details, scene splits, and question-answer (QA) pairs.

The dataset was built through the following sophisticated pipeline:

• Data filtering: English-language videos were extracted from 1.9 million videos in YouTube-Commons, and dynamic content was selected based on word density and visual dynamism.
• Automated annotation generation: Gemini 1.5 Pro was used for content selection in long videos, and GPT-4o was used to generate precise metadata for structured data output.
• Categorization: A self-built taxonomy was applied to systematically classify the video content.

FineVideo can serve as a core resource for training Video Understanding, text-based video generation (Diffusion Models), and computer vision models.