Principled conflation of major open datasets for points of interest (POIs) across the United States.
A unified, open dataset for points of interest across the United States. Built from OpenStreetMap and Overture Maps, with per-POI confidence scores calibrated against a verified validation sample so they read as probabilities that a place exists and is currently open.
Each refresh writes a new versioned folder. Inside every version:
latest/ is a server-side mirror of the most recently published version —
use it for live demos and tutorials, and pin a dated folder
(e.g. 2026-04-23-v0/) when you need a stable, reproducible reference.
Browse all versions at https://source.coop/henryspatialanalysis/openpois.
One row per real-world POI after matching OpenStreetMap features against Overture Maps places. Key columns:
The osm-parquet/ files contain the same OSM rows before conflation. This data retains the original OSM tags.
conf_mean on the conflated dataset is a calibrated probability that the place
exists and is open to the public, not a raw model score. It is produced by
mapping each POI's source score(s) through a curve estimated from an independent
validation sample of POIs whose real-world status was established by research
and human review. Curves are fitted separately for each detection pattern — in
both sources, OpenStreetMap only, Overture only — because those populations
behave very differently.
Three consequences worth knowing before you filter on it:
conf_mean_uncalibrated holds the old-style value if you
need to reproduce prior behaviour.conf_mean means something different in osm-parquet/. There it is the
OpenStreetMap turnover-model posterior — the probability the feature is still
current given its edit history — and it is not calibrated against verified
ground truth. Do not compare the two columns numerically across the two
datasets.calibration_flag records the exceptions:
Read directly from Source Cooperative's S3 mirror (no authentication):
pyarrow.fs.S3FileSystem(anonymous=True) and a
bucket-qualified path (no scheme prefix).s3:// URL plus an anonymous SECRET so its glob
expansion works over the bucket listing.Every example uses VERSION = "latest"; swap in a dated folder (e.g.
"2026-04-23-v0") when you need a reproducible pin.
The *-pmtiles/*.pmtiles archives can be loaded directly by any PMTiles
client (MapLibre + pmtiles://, OpenLayers + ol-pmtiles, etc.). See
site/ in the GitHub repo for a working example.
The snippet below is a self-contained HTML page that renders the conflated
PMTiles over MapLibre, coloured by the model's conf_mean. Save it as
openpois.html and open it in a browser — no build step, no server needed.
PMTiles are authored at zoom 14, so zoom in past z14 to see points.
The OpenPOIs dataset is released under the Open Database License (ODbL) v.1.0. Any public use must credit OpenStreetMap contributors, the Overture Maps Foundation, and OpenPOIs. Any derivative database must be shared under the same license. See https://www.openstreetmap.org/copyright and https://docs.overturemaps.org/attribution/ for upstream attribution requirements.
If you use this data in research, please cite:
Henry Spatial Analysis (2026). OpenPOIs: a unified, confidence-scored dataset of U.S. points of interest. https://openpois.henryspatialanalysis.com
Questions, bug reports, and contributions welcome via https://github.com/henryspatialanalysis/openpois/issues.
match_score, match_distance_m | Diagnostics for the OSM × Overture link |
osm_conf_mean, overture_confidence | The two raw per-source inputs to the calibration |
bbox | GeoParquet 1.1 covering struct, for spatial row-group pruning |
geohash | Within-partition sort key (precision 6) |
geometry | WKB point (EPSG:4326) |