The open databases Photerra builds on
Some of what Photerra knows about the world comes from open databases built by volunteers and scientists. This page says which ones, what we derive from them, exactly how — and how to get our derived data under the same share-alike terms.
What we derive, and from where
Photerra maintains a derived geographic database that helps our AI describe photos honestly: setting areas (parks, zoos, botanical gardens, nature reserves, protected land — polygons that tell us whether a photo was taken in a zoo or in the wild) and named features (peaks, rivers, lakes, waterfalls, bridges, monuments, lighthouses — the things a camera at a known position and heading might actually be looking at). It is extracted from OpenStreetMap, © OpenStreetMap contributors, licensed under the Open Database License (ODbL) 1.0. Because we systematically extract and transform OSM features, that derived database is a Derivative Database under ODbL, and the captions and place descriptions produced from it are Produced Works — which is why this page exists.
Alongside it we keep regional species priors — which wild animals are plausibly in an area — from GBIF occurrence records (research-grade iNaturalist observations only, restricted to records their observers published under CC0 or CC BY). Place names are cross-checked against Overture Maps and Wikidata, and archive photography comes from Wikimedia Commons with per-photo attribution (title, author, source link and license) displayed wherever the photo appears.
The method (our ODbL §4.6 disclosure)
ODbL asks that anyone publicly using a Derivative Database offer either the database itself or the method used to make it. Here is the method, in enough detail to reproduce it:
Extraction
- Queries run against the Overpass API over bounded boxes per metro or park region, in two lanes: settings (ways and relations tagged
leisure=park,leisure=nature_reserve,leisure=garden(botanical),tourism=zoo,tourism=aquarium,boundary=protected_area,boundary=national_parkand close kin) and features (nodes/ways taggednatural=peak,natural=water,waterway=river,waterway=waterfall,man_made=bridge/bridge=yes,historic=monument,man_made=lighthouse, islands, and similar view-target kinds). - Tiles that return an element-cap or a timeout are quad-split and re-fetched rather than accepted as partial; a tile that still fails is recorded as a failure in an ingest ledger, never as an empty success.
Transformation
- Relation members are stitched into closed rings (multi-way boundaries regain their interiors); rings are then decimated (vertex reduction with a small tolerance) for fast point-in-polygon tests, and stored with their bounding boxes for indexing.
- Each setting is assigned a kind with the precedence captive > cultivated > wild > urban green (a zoo inside a park is a zoo). OSM tags are trusted over titles: heritage/historic/building-tagged areas and high protect-class values are excluded from "wild"; named city parks classify as urban green even when tagged as protected areas.
- Features keep their OSM name, kind and geometry (point, box, or polyline); distances to polylines are measured to the nearest segment, and per-kind distance caps (for example, a bridge is only offered as a subject within ~1.5 km) stand in for terrain occlusion.
- Species priors: for each wild-kind setting polygon, GBIF occurrence facets are queried over the polygon (largest outer ring, counter-clockwise; bounding box as fallback), restricted to the iNaturalist research-grade dataset, animal kingdom, CC0/CC BY records. The top taxa are stored with US-English vernacular names, with regional display-name overrides (in North America, Cervus elaphus reads "Elk", never "red deer").
- Every ingest run is versioned (source release tag) and ledgered with row counts and failures, so the derived database's provenance is auditable.
Use
At caption time, a photo's coordinates (and heading, when the photo carries one) select the containing setting and the plausibly visible features; the model may only name features that pass uniqueness gates, and abstains otherwise. Captions produced this way are labelled as model-written wherever they are published — every destination gallery on this site carries that note under its photographs.
Share-alike offer. The derived database described above is available under the ODbL — email legal@photerra.com with the subject "ODbL request" and we will provide a current export, free of charge. The method above is offered under the same section (§4.6) of the license.
Related
Open data is one of two places where what we publish has consequences beyond us. The other is the coordinate itself — see where we stand on geotagging.
Attribution, everywhere it belongs
- Maps and place data: © OpenStreetMap contributors (ODbL) · Overture Maps Foundation · Wikidata (CC0).
- Species data: GBIF.org occurrence downloads; underlying records © their iNaturalist observers, CC0/CC BY.
- Archive photography: Wikimedia Commons, credited per photo (author, license, and source link) wherever it appears in Photerra.
- Map tiles on this site: served by OpenFreeMap, from OpenStreetMap data, using OpenMapTiles schema and styles.
Corrections and questions: legal@photerra.com.