SemGeo: Semantic Keywords for Cross-View Image Geo-Localization

Image geo-localization refers to determining location of images using visual information. In cross-view image geo-localization, location of images are identified by matching them against a database of geo-tagged aerial imagery. The methods based on this approach rely solely on color statistics to resolve view point changes. This has seemed to work well for large Field of View (FoV) images, but leads to poor results in low FoV setups. Images captured from standard cameras have small FoVs, these devices cannot capture images with FoV larger than 180°. Exploring information other than color statistics is important to enable geo-localization for them. We present SemGeo, a ground to aerial cross-view image geo-localization framework that considers semantic keywords from images while geo-localizing them. Experiments on standard benchmarks demonstrates that the proposed approach shows impressive gains on FoV constrained image geo-localization improving top-1% image recall rate on the CVUSA data-set from 89.18% to 94.40% and from 85.60% to 91.86% on the CVACT benchmark for 90°FoV.