Indoor 3.6M : A Dataset and Benchmark for Indoor Image Geolocation

Ogechi Blessing Onuoha, Bjørn Sand Jensen, Jan Paul Siebert, Frank E. PollickOpenReview 2026 · Submission to ICLR 2026 · Rejected · OpenReview

Abstract

Image geolocation has advanced rapidly for outdoor imagery, driven by large-scale benchmarks and strong visual cues such as landmarks, skylines, and vegetation. In contrast, indoor image geolocation remains underexplored: indoor scenes lack distinctive geographic features, are highly ambiguous, and are not adequately represented in existing datasets. We address this gap by introducing the first large-scale benchmark for indoor geolocation, consisting of \textbf{3.6 million} images across \textbf{213 countries}. We finetune state-of-the-art CLIP-based models such as Pigeon and GeoCLIP and report performance at country and continent levels using both top-$k$ accuracy as well as distance based accuracy metrics. Results highlight that continent-level geolocation is feasible, but fine grained indoor geolocation e.g street and city level geolocation remains an open challenge. This work defines a new frontier for geolocation research and provides the resources to advance it.
BibTeX
@misc{onuoha2026indoor,
  author = {Ogechi Blessing Onuoha and Bjørn Sand Jensen and Jan Paul Siebert and Frank E. Pollick},
  title = {{Indoor 3.6M : A Dataset and Benchmark for Indoor Image Geolocation}},
  year = {2026},
  url = {https://openreview.net/forum?id=Nw7vkJKHba},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2026},
}

Loading related content...

Figure