Leveraging EfficientNet and Contrastive Learning for Accurate Global-scale Location Estimation

In this paper, we address the problem of global-scale image geolocation,\nproposing a mixed classification-retrieval scheme. Unlike other methods that\nstrictly tackle the problem as a classification or retrieval task, we combine\nthe two practices in a unified solution leveraging the advantages of each\napproach with two different modules. The first leverages the EfficientNet\narchitecture to assign images to a specific geographic cell in a robust way.\nThe second introduces a new residual architecture that is trained with\ncontrastive learning to map input images to an embedding space that minimizes\nthe pairwise geodesic distance of same-location images. For the final location\nestimation, the two modules are combined with a search-within-cell scheme,\nwhere the locations of most similar images from the predicted geographic cell\nare aggregated based on a spatial clustering scheme. Our approach demonstrates\nvery competitive performance on four public datasets, achieving new\nstate-of-the-art performance in fine granularity scales, i.e., 15.0% at 1km\nrange on Im2GPS3k.\n