Shrinking AI into the edge to make video more useful
One of the other high-profile AI startups we met is OmniEyes. We talked to one of the three professor co-founders, Chun-Ting Chou, about what they are doing that’s different, success to date, target markets and future plans.
The founders have a background in wireless networking, multimedia and machine learning. They brought this together under the government’s program that funds professors to commercialize their research, and set up OmniEyes, which now has nine people working for it. The firm collects video data from dashcams and commercial-grade live street recordings to provide useful location-based information for real-time task execution.
In Taiwan, almost every car has a dashcam fitted, mainly to provide evidence under motor insurance claims. Chou said, “These collect so much information about the city, environment, gas prices, points of interest and so on, but most of the information is buried in the video. We are using AI to automatically interpret the information.”
Their intention is then to sell this ‘actionable’ information to relevant groups of users; fleet management is an initial target sector. Chou said some of the challenges for fleet managers include the cost of vehicles idling, and the high potential for traffic violations as drivers face greater pressure to deliver within ever shorter timelines. OmniEyes delivers a turnkey package for fleet managers that can help address these issues and optimize fleet operations.
We asked how what they are doing is different to the way companies like MobileEye are offering their service. Chou said that MobileEye is using data only for closed loop systems, and not collecting video data for other purposes. He added that OmniEyes’ AI engine sits in the on-board unit (OBU) in the dashcam, and fleet managers pay monthly or annual subscriptions per vehicle to get key parameters or information extracted from video images that OmniEyes collects on its platform (which collects video not just from that fleet but multiple fleets and other live street video sources).
He explained the challenge was in analyzing the videos in the OBU itself, shrinking large convolutional neural networks (CNNs) onto a small lightweight device without losing performance. “That’s what we do — we shrink the AI into the end device, and then move it to the cloud. Shrinking is the OmniEyes secret sauce.” The closed-loop system sends pseudo-labelled data from the end device to the cloud, where the data is used for training, and then shrunk, and sent back to the device OTA (over-the-air) interface.
On the issue of data privacy, Chou stressed that OmniEyes is not carrying out facial or people recognition, nor does it release road pictures. The closest they get to user-specific data is identifying license plates, but they don’t connect this back to driver’s identity. On the latter, he said they are already talking to potential customers in the U.S. interested in using the license plate detection capability.
Plugging into global networks for better chance of success
The above are a tiny selection of the startups we were introduced to in Taipei, and it certainly feels like a vibrant ecosystem that is keen to grow. The highly interconnected networks of international and local incubators and accelerators ignited by ambitious government programs to commercialize research is certainly creating a strong pipeline of hardware-based AI services companies in everything from health tech and industrial tech to agri-tech. Together with a growing army of local and foreign venture investors keen to bring some of these startups into their portfolios, this will definitely give Minister Liang-Gee Chen some of the outcomes that he’d envisioned, over the medium to long term.
【Posted by EE Times】