Minute.ly taps AI to generate highlight reels of live broadcasts

For web publishers in the video production business, there’s a lot riding on previews. Piecing together the perfect highlight reel can potentially boost click-through rates and views, but it’s a labor-intensive process — editing and curating clips takes time, skill, and a good deal of patience.

Tel Aviv, Israel-based computer vision startup Minute.ly claims to have developed a solution that’s not only scalable, but pretty much surefire. It’s a real-time software analysis tool that ingests livestreams to automatically (and instantaneously) generate attention-grabbing, heart-thumping five-to-seven-second clips from them.

You might think of it as the world’s fastest video editor, Minute.ly CEO and cofounder Amit Golan told VentureBeat in a phone interview.

“Our product is the first of its kind that monitors video in real time and is able to analyze content to provide key moments and promote those moments instantaneously,” Golan said. “The initial idea was ‘How can we automatically edit videos and find the most interesting parts?‘ We’re trying to find ways to increase the chance that users will watch the video.”

An algorithm that knows which scenes are objectively more compelling than others might sound like science fiction, but the secret is data, Golan said. Lots of data. Much of it was crowdsourced from a Tinder-like Android app the Minute.ly team launched in 2013; users swiped right on auto-generated video snippets that caught their eye.

“We do a lot of analysis in the background to collect and understand the peak moments in each video,” he explained, “combined with AI on the video processing side to achieve close to 100 percent accuracy.”

The resulting platform can boost a livestream audience by 13 percent and advertising revenue by up to 37 percent. It’s already been used by “top broadcasters” for the 2018 World Cup Russia, Golan said, and by “dozens” of sports, news, and entertainment websites.

“Through our first of its kind deep video analysis and crowdsourcing technology, we automatically discover what pieces of content truly engage users through real-time personalized experience,” added Golan. “We can identify and create the most engaging previews to drive user engagement.”

Minute.ly isn’t the first company to tap machine learning for video preview and thumbnail generation. Google uses deep learning — a subfield of artificial intelligence that employs brain-like neural networks in identifying data patterns — to power YouTube’s automated thumbnail selector. (A point worth mentioning: YouTube, responding to criticism following an experiment in which 0.3 percent of viewers saw auto-generated thumbnails rather than custom-created images, said in June that it has “no current plans to change or remove the ability to add custom thumbnails.”) And researchers at Netflix have reportedly experimented with AI that creates shows and movie trailers tailored to subscribers’ preferences.

But Golan insists he’s not trying to take human editors or creators out of the loop. Instead, he explained, the goal is to help publishers make the most of their video inventory.

“By 2020, streaming video will contribute to 80 percent of online traffic … That’s a lot of competition, and most publishers won’t generate enough video revenue to cover expenses,” he said. “Our product helps their chances.”

Minute.ly’s new real-time preview generator joins other AI-powered tools in its product suite, including Top Videos, which automatically aggregates publishers’ top-performing video articles and presents recommendations to viewers, and Smart Video Preview, which generates teasers from existing videos. They’re available through Control Panel, Minutely’s bespoke analytics dashboard that displays real-time stats like click-through rates and video engagement.