OSM-Europe-1k

A 1,000-image street-level geolocation benchmark for Europe, sampled from the OpenStreetView-5M (OSV-5M) test split. Intended as a contamination-free, openly-licensed reference set for evaluating image→GPS models. Companion benchmark to a master's thesis at FH JOANNEUM (Florian Leber, 2026).

What's in this repo:

• The 1,000 OSM/Mapillary image bytes are bundled directly under images/ (66 MB) — re-distribution is allowed by the upstream CC-BY-SA 4.0 license, with attribution to Mapillary contributors.
• The GSV image bytes are not included (Google's Street View Static API TOS prohibits redistribution). You fetch them locally via download_gsv.py using your own API key.
• Coordinates for both splits (and the OSM∩GSV intersection) ship as CSV manifests. download.py is kept for users who prefer regenerating the OSM split from upstream OSV-5M.

Coordinates	1,000 (OSM) / 1,000 (GSV-attempted) / 837 (paired OSM ∩ GSV)
Region	Europe (lat 34..72, lon −25..45)
Source	osv5m/osv5m, test split
Image source (OSM split)	Mapillary street-level (via OSV-5M)
Image source (GSV split)	Google Street View Static API (4 headings × pano)
Sampling	streaming-shuffle, seed=42, deterministic
License	CC-BY-SA 4.0 (matches OSV-5M / Mapillary attribution requirement)

Files

File	Rows	Schema	Notes
images/	1 000	00000.jpg…00999.jpg (~66 MB total)	OSM/Mapillary image bytes, redistributed under CC-BY-SA 4.0. Filenames match metadata.csv.
metadata.csv	1 000	filename,lat,lon,country,osv5m_id	OSM/Mapillary subset. filename matches the bundled images/ files.
metadata_gsv.csv	1 000	filename,lat,lon,country,osv5m_id,headings_ok,n_ok	GSV-paired subset. headings_ok is a |-separated list of headings (0/90/180/270) for which Google returned a real (non-grey) panorama; n_ok is its count. Locations with n_ok = 0 are absent here.
metadata_paired.csv	837	filename_base,lat,lon,country,osv5m_id	Strict intersection: locations where the OSM image and all 4 GSV headings exist. Use this split for like-for-like model comparisons across image sources.
download.py	–	–	Reproduces metadata.csv + images/ from the upstream OSV-5M test split.
download_gsv.py	–	–	Fetches GSV imagery at the same coordinates (requires GOOGLE_STREETVIEW_API_KEY).
eval.py	–	–	Standalone evaluator; takes a predictions CSV/JSON, reports im2gps-style threshold accuracy + great-circle error.

Why "OSM"?

OSV-5M pairs Mapillary street-level imagery with the OpenStreetMap-affiliated open-imagery ecosystem. It is distinct from the closed Google Street View collections used elsewhere in geolocation literature, and so offers a cleanly licensed, reproducible out-of-distribution probe for thesis-grade geolocation models trained on private GSV data.

Why the test split?

Many published image-geolocation models — including OSV-5M's own baselines — are trained on the OSV-5M train split. Sampling our benchmark from test guarantees that none of these coordinates have been seen during such training, so reported numbers reflect generalisation rather than memorisation.

Quickstart

pip install huggingface_hub pandas pillow

# 1. Pull this dataset (manifest + 1,000 OSM/Mapillary images, ~67 MB)
huggingface-cli download lebfla11/osm-europe-1k --repo-type dataset --local-dir osm_eu_1k

# 2. (optional) Regenerate the OSM split from upstream OSV-5M instead
#    (~80 MB CSV + 2.5 GB zips, cached by HF — only needed if you want
#    to confirm the seed=42 sampling or change `--target`):
python osm_eu_1k/download.py --target 1000 --seed 42

# 3. Fetch the GSV-paired images yourself (requires Google API key + ~$2
#    of paid quota at the time of writing). Not bundled, per Google TOS.
GOOGLE_STREETVIEW_API_KEY=... python osm_eu_1k/download_gsv.py

Evaluation

metadata.csv and metadata_paired.csv are schema-compatible with im2gps_europe.csv-style geolocation pipelines. The included eval.py is self-contained (no project imports):

python osm_eu_1k/eval.py --predictions path/to/predictions.json

Predictions format

CSV:

filename,pred_lat,pred_lon
00000.jpg,48.85,2.35
00001.jpg,52.51,13.40
...

…or JSON (a list of objects with the same keys; extra fields are ignored).

Reported metrics

• mean_dist_km, median_dist_km — great-circle error
• within_{1, 25, 200, 750, 2500}_km — accuracy at the standard im2gps thresholds

Reference numbers

A Street-View-trained classifier (StreetCLIP encoder + linear head over 8,494 H3-res-5 cells, last-2 transformer blocks fine-tuned for 1 epoch on raw images) reaches the following on the paired 4-view GSV split (i.e. GSV imagery the model has never seen, predicted by 4-view feature mean fusion + top-1 cell decoding):

split	top-1 (cell)	within 25 km	within 200 km	mean km	median km
metadata_paired (GSV, 4-view, n=837)	3.6 %	4.7 %	33.8 %	523	345

These are the numbers any reasonable model should be able to beat — the cell-classifier hasn't been trained on Mapillary or any GSV slice from OSV-5M, so it underperforms by design. Use them as a sanity check that your eval pipeline is wired up the same way (4 views, top-1 cell, mean-of-features fusion).

License & attribution

The CSV manifests in this repo are released under CC-BY-SA 4.0 to match the upstream OSV-5M / Mapillary attribution requirement. If you redistribute them or any derivative product (e.g. cached extracted images), you must:

1. Credit Mapillary contributors and OSV-5M (citation block below);
2. Keep the same CC-BY-SA 4.0 license on derivatives.

The accompanying Python scripts (download.py, download_gsv.py, eval.py) are MIT-licensed.

The bundled OSM image bytes under images/ are redistributed under CC-BY-SA 4.0 as inherited from OSV-5M and Mapillary. If you fork or re-publish them, you must keep CC-BY-SA 4.0 and credit Mapillary contributors.

The GSV image bytes are governed by the Google Maps Platform Terms of Service and must not be redistributed by anyone, including via this repo or any fork — that's why only the coordinates + a fetcher script are published for the GSV split.

Citation

@inproceedings{astruc2024osv5m,
  title     = {{OpenStreetView-5M}: The Many Roads to Global Visual Geolocation},
  author    = {Astruc, Guillaume and Dufour, Nicolas and Siglidis, Ioannis and Aronssohn, Constantin and Bouia, Nacim and Fu, Stephanie and Loiseau, Romain and Nguyen, Van Nguyen and Raude, Charles and Vincent, Elliot and Xu, Lintao and Zablocki, Eloi and Landrieu, Loic},
  booktitle = {CVPR},
  year      = {2024}
}

@mastersthesis{leber2026streetview,
  title  = {Street View Image Geolocation in Europe via Geocell Classification},
  author = {Florian Leber},
  school = {FH JOANNEUM University of Applied Sciences},
  year   = {2026}
}

Contact

Florian Leber — florian.leber@edu.fh-joanneum.at Master's thesis at FH JOANNEUM University of Applied Sciences. Companion demo: lebfla11/streetview-eu-geocell-demo