Computer Vision News - September 2018

Computer Vision News Application 14 Application: ZOOX drive on the road autonomously using sensors: the more the system matures, the more sophisticated and hard it gets to improve it . You fix things, but at the same time you could make other things worse. It's very hard to make sure you consistently improve and narrow in on the problem. It took a reasonable amount of training effort to actually convert the system to a point where it improved over time without exploding with complexity. Then they began building solutions that didn’t require too much effort to maintain and still had the potential to scale up. He elaborates: “ You may need to understand more of how we develop. We have a certain regime that is called our ODD (Operational Design Domain) , where we define the types of situations we intend to handle. We removed a lot of situations out at the beginning. Otherwise, we would be drowning in complexity . Let’s use an example. Initially, we took out left turns and other scenarios as well. When we start driving, and we get better, we bring more of these things in because, at the end of the day, we have to solve them. So the system gets more complex. The scenarios that we have to handle get more complex. At the same time, you want to have a solution that you can maintain, you can grow, that doesn't break down once you change out a sensor, use a different type of sensor, change sensor placements, or cameras… This is something I find very challenging. Also, it’s very challenging once you change something to validate that nothing breaks. At the same time, you need to provide solutions that are able to run on the vehicle. You can have the best algorithm in theory, but if you can’t compute it in real-time, it doesn't help you. ” David comments on the public misconceptions of this technology: “ People talk about vehicles on the road today that have driver assistance. They take that technology as an autonomous system, but it's not. It’s an assisting system. To switch to an autonomous system , there is much, much more to it. An assisting system is aimed mainly on emergency brakes, and you only need to make sure that it doesn't misfire. It never guarantees the safety of a pedestrian. Whereas, we have to do both. It can't misfire. It should never break for no reason, but if there is a pedestrian, it must always break. This a completely different problem, and we take safety very seriously as a company. ” Sarah expounds: “ What David says is very important: people need to realize the massive difference between an L3 system and an L4 or L5 system , and that the sort of issues that these systems might have are totally different. L3 systems are designed to give the human “We take safety very seriously as a company”