Startup Matroid uses AI to pluck images from streams of video

If you’re a brand advertiser at a car company or a cereal maker, you may want to know when your product appears, unscripted, in hundreds of hours of TV shows or online videos. How to track that without watching all of it? An artificial intelligence startup has the answer.

Matroid, founded by Stanford University adjunct professor Reza Zadeh, can scan video for all sorts of things like when President Donald Trump appears, or how often a man holding a gun is recorded. Users can easily write a filter—Matroid calls them detectors—of their own to find particular people or objects, or they can pick from a library of pre-programmed filters designed by the startup.

Matroid is focusing initially on customers who want to analyze television appearances or scan surveillance video. In the first case, a firm may want to track which political candidates got more TV time or which brand of car appears more often on a particular show or network. One company talking to Matroid wants to post cameras near outdoor billboards and track passersbys’ reactions to the ads. The startup also is talking to makers of surveillance cameras, said Zadeh, who declined to name them.

For surveillance, think of the rapidly increasing amount of video from police body and dashboard cams, home security systems and municipal closed-circuit cameras. Say a city police department wants to track how officers respond when they see a man holding a gun, Matroid’s tools will help the department search all the archived camera footage for men with guns. As public requests increase for such police footage, officials need easier ways to find the faces and personal details that have to be blurred before camera video can be released publicly.

The company, founded early in 2016 and funded by New Enterprise Associates Inc., unveiled its product Saturday at the Scaled Machine Learning conference at Stanford. Zadeh declined to specify the amount of funding they’ve received.

Google, Facebook and Microsoft, among other big firms, employ computer vision scientists to determine what or who is in a particular photo. But most companies don’t have the kind of AI firepower to do it themselves. With the explosion of video from YouTube to police bodycams, Matroid is seeking to fill the need for rapid scanning of massive volumes of moving images.

“Google can give you pictures of cats, but not cat with grandpa or cat with grandpa and Christmas tree or with your son,” said Pete Sonsini, a general partner at NEA. “It’s really powerful for any human to be able to create a detector that can identify any image or set of images or face from their dataset.”

Here’s how it works: users select pictures or parts of pictures that show what they want to find. Matroid’s algorithms learn from those images and create a detector that can search for the same thing in videos or other photos.

Zadeh is hoping to benefit from the expertise of attendees at the conference and others who want to create detectors. The 600 event attendees will get to test the product and be asked to give feedback to help improve it. Detectors will be publicly available for others to use and those who designed them will get a cut of the money made by Matroid.

To grow, Matroid will have to decide where to focus because there are many potential uses, Sonsini said. Plus, they’re not the only company working in the broader area.

“Some level of image recognition is not a novel idea,” he said. “But how they are doing it is novel.”