Computer Vision News - June 2019

in the same world that we operate in so self-driving cars or any type of autonomous robot. Again, the same thing: if you sell a car for a lot of money, but you can’t drive it at night because our algorithms do not work at night, I would not pay for it. So you need this type of robustness! ”, he continues. Existing localization algorithms only go so far and haven’t yet solved the problem of long-term localization under a wide range of conditions. “ You try to match scenes with other random scenes, but you don’t know what is inside the scenery. You don’t have any representation of the semantics of the scene. Now, with neural networks and all of these recent ideas, we can build on top of that and actually understand what is going on in the scene ” says Vassileios. Long-term visual localization has many applications in industry, and the organizers of this challenge hope the workshop will motivate others who might have a better solution. Vassileios shares: “ The obvious one is augmented reality. The basic rule is if you want to augment reality then you need to understand where you are. Self-driving cars is another very important aspect of it. ” In addition, visual localization can allow extra-precise delivery: a drone system that can deliver a package directly to someone’s balcony or other precise, private location. For example, the winning entry in Scape Technologies' first Hackathon (team Inition) demoed smartphones in which users draw a square with their finger inside the camera view. Then the drone lands exactly in this area. “ Anything that has to do with extremely accurate position systems can benefit from this work. We use GPS systems now. We can use vision-related technology in the next years. Hopefully, we will have much, much better results! ” Vassileios elaborates. The challenge aims to capture the impact of changes on localization algorithms, to create interesting, realistic datasets and apply state-of- the-art algorithms to the models. Looking at some of the immediate results, they still remain a long way from solving some of the datasets. Current visual localization methods focus mostly on very accurate matching-based localization, which uses an image and then matches the image within the accuracy of centimeters. Another aspect involves deriving a very rough estimation, very 30 Computer Vision News Challenge and Workshop - CVPR 2019 Challenge