Computer Vision News - May 2019

neural networks. Antonio shares: “ There are other approaches called end- to-end driving. You drive your car. You collect perception data, like images, for instance. Together, with these images, you link some information from the car: the speed, the acceleration, the breaking. All of these measurements are acting as the ground truth; the perception data and this ground truth are used to train the AI driver .” “ In this case, you need very few manual interventions to train the networks, at least for annotating the data. You don’t need to explicitly ask the network to detect anything like pedestrians, cars, or whatever is around you. You don’t have this explicitly in the system. The system learns what is important and what is not important for driving. These are two very different paradigms ” he adds. For the challenge, some teams follow the first paradigm, while others follow the latter. The results will reveal how these paradigms compete and what makes one better or worse than the other. The challenge has four tracks, each with different difficulties. In one track, the team uses different sensors. In another, the team receives all the information so that people have a perfect perception, and they only need to focus on control. Antonio elaborates: “ A novelty coming up now is more cities showing more complex scenarios like multiple lanes, with their respective individual traffic lights and roundabouts. Then cars have to negotiate intersections with other cars. They have to do maneuvers like overtaking another car. They have to take notice of pedestrians walking around. Also, from the very beginning of CARLA, we focused a lot on bringing variability. In CARLA, you have different light conditions and different weather conditions like rain… It’s really challenging because there are a lot of variabilities that you can simulate .” Antonio has his own PhD students working on CARLA. So far, they made many unexpected discoveries, things they wouldn’t have learned by processing a static dataset. He looks forward to seeing the end results: “ I think the challenge is going to be great: we will see how CARLA can help the community to bring new knowledge about autonomous driving in different paradigms. ” Pauline Luc 35 Computer Vision News Feedback of the Month I had a pleasure working with RSIP Vision’s team and personally with Ron on our highly sophisticated image guided project. RSIP Vision demonstrated both professionality and concern during the product development - the right team to make it happen! Sasi Solomon Topspin CARLA Autonomous Driving Challenge