We scanned 9,487,758 public geotagged photographs from 246 countries and asked one question: when the world picks up a camera, what does it point at?
Each country is painted by the total public photos we found in it (log scale). Hover to see its most-photographed thing — cities, regions, and country names filtered out. Click for the full breakdown.
Nine things we didn't know until we ran this pipeline.
YFCC100M (OpenAI subset) — 15M Flickr uploads with public Creative Commons licensing, pre-sharded into 4,094 metadata files on HuggingFace. We filtered to the 9.49M rows with latitude + longitude.
Burla spun up a 13-node cluster (80 vCPUs each). Peak concurrency: 967 simultaneous Python workers — each processing one shard of ~2,300 photos and reverse-geocoding every lat/lon into a (country, admin, city) tuple against a 30k-city KD-tree.
Three passes, all on Burla: extract (118 s · 967 workers), tokenize (107 s · 640 workers), and reduce (230 s · 64 workers). End-to-end under 8 minutes of wall-clock.
We scored every user-tagged phrase with TF-IDF across all 246 countries, then filtered out cities, regions, and country-name aliases so "Paris" wouldn't be France's most photographed thing. What's left is landmarks, events, subcultures, and obsessions.
Every number you see was computed from the raw data — no editorial curation. Phrases come straight from the authors' Flickr tags. See the Burla demo code.