LocDiffusion: Identifying Locations on Earth by Diffusing in the Hilbert Space
Journal:
arXiv
Published Date:
Mar 23, 2025
Abstract
Image geolocalization is a fundamental yet challenging task, aiming at
inferring the geolocation on Earth where an image is taken. Existing methods
approach it either via grid-based classification or via image retrieval. Their
performance significantly suffers when the spatial distribution of test images
does not align with such choices. To address these limitations, we propose to
leverage diffusion as a mechanism for image geolocalization. To avoid the
problematic manifold reprojection step in diffusion, we developed a novel
spherical positional encoding-decoding framework, which encodes points on a
spherical surface (e.g., geolocations on Earth) into a Hilbert space of
Spherical Harmonics coefficients and decodes points (geolocations) by
mode-seeking. We call this type of position encoding Spherical Harmonics Dirac
Delta (SHDD) Representation. We also propose a novel SirenNet-based
architecture called CS-UNet to learn the conditional backward process in the
latent SHDD space by minimizing a latent KL-divergence loss. We train a
conditional latent diffusion model called LocDiffusion that generates
geolocations under the guidance of images -- to the best of our knowledge, the
first generative model for image geolocalization by diffusing geolocation
information in a hidden location embedding space. We evaluate our method
against SOTA image geolocalization baselines. LocDiffusion achieves competitive
geolocalization performance and demonstrates significantly stronger
generalizability to unseen geolocations.
