At Disrupt NY 2017, Deep Science AI made its debut on stage, showing in a live demonstration how its computer vision system could spot a gun or mask in CCTV footage, potentially alerting a store or security provider to an imminent crime. The company has now been acquired through a friendly merger with Defendry, which seeks to more widely deploy the tech.
It’s a great example of a tech – focused company seeking market access, and a market – focused company seeking the right tech.
The idea was that if you have a chain of 20 stores and 3 cameras at each store, and people can only reliably keep an eye on 8 – 10 feeds at a time, you’re looking at a significant investment in staff just to make sure they’re not pointless. If instead you used the middle layer of Deep Science AI that highlighted shady situations such as drawn guns, one person might possibly keep an eye on hundreds of feeds. While they didn’t take the cup that year, it was a good pitch.
“The battlefield of TechCrunch was a great starting point to get our name and capabilities out there,” Deep Science AI co – founder Sean Huver said in an interview (thanks for the plug, Sean). “In the retail space request pilots, we had some really big names. But we quickly discovered that, as far as what actually happens next, there was not enough in the infrastructure.”
That is, things like automated security dispatch, integration with servers and hardware from the private company, that kind of thing.
“The monitoring around the AI technology really needs to be built rather than the other way around,” Huver admitted.
Meanwhile, Pat Sullivan at Defendry was working on setting up automated workflows for the internet of things devices — from adjusting the A / C if the temperature exceeds certain limits, to notifying a company of serious problems such as robberies and fires at some point.
“One of the most important alerts that could happen is someone who has a gun and does something wrong,” he said. “Why can our workflows not activate the alert with notifications and tasks, etc? That led me to look for a dataset of weapons and dangerous situations that led me to Sean.
Although the company was still only in the prototype phase when it was on stage, the success of its live demo with a team member setting off an alert in a live feed (good to try) indicated it was actually functional — unlike many other companies advertising the same thing, as Sullivan discovered.
“Everyone said they had the goods, but they really didn’t,” he said, when you evaluated them. “For us, nearly everyone wanted to build it — for a million dollars. But we were thrilled to see when we came across Deep Science that they could actually do what they said they could do.”
Ideally, he went on to suggest that the system might not only be an indicator of ongoing crimes, but crimes that are about to begin: for example, a person wearing a mask or pulling a gun out of a parking lot could trigger outdoor doors to lock. And when a human checks in, either the police could be on their way before the person reached the entrance, or it could be a false positive and the door could be unlocked before anyone even noticed anything had happened.
Now, the bias in computer vision algorithms is one part of the equation that is mercifully not necessarily relevant here. We’ve seen how women and colored people — to begin with — are disproportionately affected by error, misidentification, etc. I asked Huver and Sullivan if they had to accommodate these problems.
Fortunately, they explained, this tech is not relying on facial analysis or anything like that.
“We’re really pushing this issue around because we’re focusing on very distinct objects,” Huver said. “There is an analysis of behavior and motion, but the accuracy rates are simply not there for what we need to perform on a scale.”
“We don’t keep a list of criminals or terrorists and try to match the list with their face,” Sullivan added.
The two companies talked about licensing but ultimately decided that they would work best as a single organization and finalized the paperwork just a few weeks ago. They refused to detail the finances, which is understandable given the hysteria surrounding start – ups and valuations of AI.
They work with Avinet, a security hardware provider who will be the preferred vendor for customer set-ups put together by the Defendry team and has invested an undisclosed amount in the partnership. We will closely monitor the progress of this success story on the Battlefield.