FineVideo: 43K Annotated Video Dataset - Filtering & Annotation Pipeline

AI Impact Summary

FineVideo's development centers around addressing the scarcity of open-source video datasets, leading to the creation of a 43k-video dataset annotated with rich metadata. The process involves filtering YouTube-Commons, applying dynamic content filters based on word density and visual dynamism (using FFMPEG), and then categorizing the remaining videos using GPT4-o and Llama 3.1. This multi-stage pipeline highlights the team's innovative approach to data acquisition and preparation for training video AI models, particularly for tasks like video generation and computer vision.