LAION drops massive open video dataset with 10 million hours of footage for AI research
decoder ·
LAION has released the Big Video Dataset (BVD), described as one of the largest open video datasets for AI research. The dataset originates from 1.3 billion video URLs found in CommonCrawl, from which the team downloaded 80 million videos totaling 10 million hours. From these, they extracted 55 million clips with automatically generated video and audio descriptions, along with 300 million still images. Most of the videos come from YouTube and are predominantly in English. According to the accompanying paper, models trained on BVD outperform models trained on InternVid by up to 2.1 percentage points on common video-to-text benchmarks. The training method links video, audio, and text to learn which visual content matches descriptions or sounds. LAION is releasing the dataset for research purposes only, citing a 2024 Hamburg Regional Court ruling that permits collecting copyrighted content for non-commercial research, and asks users to respect creators' rights. The dataset and code are freely available.
decoder