Google Launches Video Analytics Challenge
It’s time to organize all those YouTube cat videos.
A competition designed to develop classification algorithms for labeling millions of YouTube videos is designed to sort out Google’s video trove by creating new search tools for archived footage.
The search giant and YouTube owner (NASDAQ: GOOGL) announced the video classification challenge this week that includes cash prizes up to $30,000. Video labels would be assigned using the YouTube-8M dataset released last fall and created from more than 7 million YouTube video URLs, which works out to 450,000 hours of video. Those labels are drawn from more than 4,700 classification tags, or about three search labels per video.
“Video represents another great opportunity to detect and recognize objects and understand human actions and interactions with the world,” Google noted in a blog post announcing the competition. “Improving our understanding of video imagery can lead to better video search, organization and discovery—for personal memories, enterprise video archives or public video collections.”
Along with the YouTube dataset, competitors can use Google Cloud Machine Learning, the TensorFlow machine learning library or other machine learning frameworks. (Google released the 1.0 version of TensorFlow this week.)
As machine learning is applied to the organization of unstructured data such as video, the company noted that many advances in machine “perception” have come through large labeled datasets. That in turn has accelerated research in parsing images. YouTube-8M is among a list of image-related datasets released by the search giant.
Google software engineers note that the updated video dataset also includes “pre-computed” audio features. That, they assert, opens up opportunities for “new research on joint audio-visual (temporal) modeling.”
“We designed this dataset with scale and diversity in mind, and hope lessons learned here will generalize to many video domains,” noted Paul Natsev, a Google software engineer, adding YouTube-8M captures more than 20 diverse video domains. “The challenge can also accelerate research by enabling researchers without access to big data or compute clusters to explore and innovate at previously unprecedented scale.”
The Google challenge reflects the need to organize and classify the huge amounts of video data being uploaded to platforms such as YouTube on a daily basis. With daily uploads growing exponentially, Google, Facebook (NASDAQ: FB) and others are developing and deploying analytics tools to make sense of it all.
“Video is the fastest growing data type on the Internet,” Brad Hunstable told Datanami last May after IBM (NYSE: IBM) acquired his former company, Ustream. “It’s also the most emotional and arguably one of the most powerful mediums on the Internet.”
The results of the Google video classification challenge will be announced during a workshop scheduled as part of an IEEE computer vision conference in July.
The Google challenge includes awards totaling $100,000. Cash prize details are here.
Analyzing Video, the Biggest Data of Them All
MIT Uses Video to Train Machine Vision System