INDOOR-3.6M : A Multi-Modal Image Dataset for Indoor Geolocation

Ogechi Blessing Onuoha, Bjørn Sand Jensen, Jan Paul Siebert, Frank E. PollickOpenReview 2025 · Submission to ICLR 2025 · Withdrawn · OpenReview

Its authors withdrew this paper, so there is no full text to read or summarize.

Abstract

Indoor image geolocation, the task of determining the location of an indoor scene based on visual content, presents unique challenges due to the constrained and repetitive nature of indoor spaces. Current geolocation methods, while advanced in outdoor contexts, struggle to perform accurately in indoor environments due to the lack of diverse and representative indoor datasets. To address this gap, we in- troduce INDOOR-3.6M, a large-scale dataset of geotagged indoor imagery span- ning various residential, commercial, and public spaces from around the world. In addition to the dataset, we propose a new sampling methodology to ensure ge- ographic diversity and balance. We also introduce INDOOR-15K, a benchmark for evaluating indoor-specific geolocation models. Finally, we demonstrate the dataset’s utility by finetuning GeoCLIP using our dataset, which shows significant improvements over the GeoCLIP baseline on our test set and other benchmark test sets.
BibTeX
@misc{onuoha2025indoor,
  author = {Ogechi Blessing Onuoha and Bjørn Sand Jensen and Jan Paul Siebert and Frank E. Pollick},
  title = {{INDOOR-3.6M : A Multi-Modal Image Dataset for Indoor Geolocation}},
  year = {2025},
  url = {https://openreview.net/forum?id=BQfAqi3Xq3},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2025},
}

Loading related content...

Figure