Project link: https://research.google.com/youtube8m/

YouTube-8M is a large-scale video dataset from Google with 7 million video URLs over 4,716 classes and 450,000 hours of video. Google also provides pre-computed, state-of-the-art audio-visual features, so that one can build their model based on these features with ease. Training from raw videos may take weeks, which is not reasonable in normal situations. This dataset's goal is to achieve video understanding, representation learning, noisy data modeling, transfer learning, and domain adaptation for videos.