Computer Vision News - June 2019

29 Long-Term Visual Localization under … addressing in the sense that it goes in the direction of what makes a place a place. It’s a very common problem in computer vision, or in the general understanding of how humans work as well. ” What makes a place a place : that’s what it boils down to, really. Humans can instinctively understand such cues. The workshop hopes to develop machine learning techniques to do the same. "We want to figure out how do I reliably recognize that I am at a certain position while the scene changes. To me, that is the essence of localization ”, Torsten explains. Current long-term visual localization algorithms depend on a scene representation constructed from images and must be robust enough to capture all potential viewing conditions including under different illumination conditions, seasonal conditions, and other changes over time. The dominant method involves recording images of a scene, which will remain valid forever. This works fine without much localization. However, any number of changing conditions can dramatically affect the appearance of a scene such as from day to night, from indoor to outdoor. When looking at outdoor scenes, vegetation grows. Buildings get renovated and demolished. Torsten explains: “ If you want to augment reality, you need to know where you are in the world in order to be able to augment the view of the user with 3D objects that are in the scene. We would like to do this as accurately as possible, at pixel level or even more accurate, and we would like to do this robustly. If you can only do it on sunny days between 12 and 2 PM, who’s going to use this, right? ” “ Similarly, you need this for any type of intelligence system that needs to operate Challenge Computer Vision News Changing conditions can dramatically affect the appearance of a scene