Computer Vision News - April 2021

349 Iris Automation platform to give accurate and precise estimations of where things are in space. “I am personally against the ideas of using deep learning to completely replace certain things that could be modelled in a general but accurate way by geometry,” he tells us. “Geometry models the world so well, why completely replace it and try to learn things end to end? I understand the motivation, but from a practical point of view I think it’s better not to do it.” The following video is a presentation of a live test scenario of the Iris Automation Collision Avoidance System. “You’re trying to detect a moving object that is very far away and pretty small in the image,” Alejandro explains. “If you give that problem to a recent graduate, they will probably throw deep learning at it and what you end up with are some good detections but a ton of false positives. The way deep learning operates right now is not going to solve this problem. The number of false positives would be too high.” But why are false positives such a bad thing? If there is an obstacle, it will be avoided, and if there isn’t, won’t it be a case of no harm, no foul? “Imagine if that is happening 10 times per minute,” Alejandro counters. “All those triggers will end up executing an avoidance manoeuvre. This could be an aggressive manoeuvre and if you’re doing that all the time then your missions are totally disrupted. It might mean that a manual pilot needs to take over.” “Potential economic savings are huge, but they can also save lives…” “The way deep learning operates right now is not going to solve this problem. The number of false positives would be too high!” A Cessna 162 aircraft makes three passes across the drone’s field of view (FOW)