In the same vein as ImageNet — a massive trove of categorized images utilized by researchers to improve machine learning, image understanding, and search functionalities across the web — Google says that it is now looking to create a similar database for videos.
On its Research Blog yesterday, the search giant announced the release of YouTube-8M, a database of 8 million YouTube videos — comprising over 500,000 hours of video — that can be used for video analysis, and thereby to help improve video search and discovery across the web. Google says the YouTube-8M — comprised exclusively of public videos with over 1,000 views — has been categorized into 24 top verticals (including Arts & Entertainment, Games, Autos & Vehicles, and Sports), as well as 4,800 more specific classes and 1.8 million total labels.
Prior to the YouTube-8M, the largest such video database was the Sports-1M, Google says, comprising 1 million sports-related YouTube videos categorized into 500 different classes. The fact that video is far more time-consuming to annotate manually than images and much more expensive to process and store makes the YouTube-8M a substantial accomplishment.
“We believe this dataset can significantly accelerate research on video understanding,” Google software engineers Sudheendra Vijayanarasimhan and Paul Natsev wrote in the post, “as it enables researchers and students without access to big data or big machines to do their research at previously unprecedented scale.” The YouTube-8M is explorable in browser form right here (check out a screencap below), and you can read more about the project in a technical report submitted by researchers to the Cornell University Library.