ICLR 2026  ·  International Conference on Learning Representations

HierLoc: Hyperbolic Entity Embeddings
for Hierarchical Visual Geolocation

1 Huawei Riemann Lab   ·   2 Technical University of Munich
0% Error Reduction
0% City Accuracy ↑
Inference Speedup
0k Entity Embeddings
Overview

HierLoc in 60 Seconds


Abstract

A New Geometry
for the Globe


Visual geolocalization — predicting where an image was taken — remains challenging due to global scale, visual ambiguity, and the inherently hierarchical structure of geography. Existing paradigms rely on either large-scale retrieval, which requires storing a large number of image embeddings, grid-based classifiers that ignore geographic continuity, or generative models that struggle with fine detail.

We introduce an entity-centric formulation of geolocation that replaces image-to-image retrieval with a compact hierarchy of geographic entities embedded in Hyperbolic space. Images are aligned to country, region, subregion, and city entities through Geo-Weighted Hyperbolic contrastive learning, directly incorporating haversine distance into the contrastive objective.

This hierarchical design enables interpretable predictions and efficient inference with 240k entity embeddings instead of over 5 million image embeddings on the OSV5M benchmark, reducing mean geodesic error by 19.5% while improving fine-grained subregion accuracy by 43%.

Explainer

Methodology Explained


Citation

BibTeX


BibTeX

@misc{gadi2026hierlochyperbolicentityembeddings,
      title={HierLoc: Hyperbolic Entity Embeddings for Hierarchical Visual Geolocation},
      author={Hari Krishna Gadi and Daniel Matos and Hongyi Luo and Lu Liu
              and Yongliang Wang and Yanfeng Zhang and Liqiu Meng},
      year={2026},
      eprint={2601.23064},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.23064},
}