The challenge, which comes with a $30,000 prize for the first-place finisher (and $25,000, $20,000, $15,000 and $10,000 for the next four teams), asks developers to classify and tag videos from Google’s updated YouTube-8M V2 data set. This data set features a total of 7 million YouTube videos that add up to 450,000 hours of video. YouTube-8M already includes labels, too, and developers can use this as their training data. The challenge then is to tag 700,000 previously unseen videos.
The company is launching this new challenge on the same day of TensorFlow’s 1.0 release — and that’s probably no coincidence. Still, Google doesn’t restrict developers to using its own machine learning framework. They can use other frameworks, too. Given that the full frame-level data set is 1.71 TB large, though, and lives on Google’s Cloud Platform, chances are most developers will want to use Google’s own services to train their models (and they can get a few extra free credits to use the Cloud Platform, too, once they run out of their free allotment).
Just last week, Google also launched the YouTube-BoundingBoxes data set. As the name implies, this data set consists of bounding boxes (for a total of 5 million videos) that track objects across frames. This isn’t something the developers will use in this new challenge, but it clearly shows Google’s interest in video classification. YouTube, after all, keeps growing, and probably sees more searches than most of Google’s search engine competitors.