Computer Vision News - July 2021

Aysim Toker 27 Best of CVPR 2021 “ When we try to geo-localize a given street view, there is a huge domain gap between the satellite and street image, ” Aysim explains. “ We propose to solve this by taking a satellite image, performing some simple mathematical transformations, and then synthesizing a content-preserving street-view image using generative adversarial models – basically, conditional GAN. ” She says that designing the architecture was the really challenging part. They didn’t know at first that when synthesizing a realistic image and geo-localizing the same image, the two aspects of the learning procedure interact and reinforce each other. “ This mutual reinforcement was really exciting because it was done in one single architecture, ” Aysim tells us. “ I think it will encourage other people to learn multiple things by fusing some tasks. ” One major application of this work in the real world could be to improve the accuracy of GPS. As we all know very well, the GPS we use in our cars every day is good, but it is not always accurate enough. Thinking about immediate next steps, Aysim says they are currently working on a solution for finding the orientation of an image if the street view is not oriented. We ask Aysim if anything ever went wrong – did the model ever find itself in completely the wrong place?