Geolocating an Unnamed Island from One Photo with Geometry and a CUDA Brute-Force Search

JeremyCMorgan · x · 2026-09-03

Security researcher yassa9 published a writeup for the gralhix 004 challenge: identify a resort, its island coordinates, and the camera's cardinal direction from a single aerial photo. Rather than “cheating” with Google Lens, the author solved it with math and programming.\n\nApproach:\n- With EXIF metadata useless, extracted three landmarks (the target islet and two neighboring islands) into a relative-geometry fingerprint\n- Filtered global coastline data by a tropical-latitude bounding box, applied local density filtering and clustering, then generated landmark triplets\n- Offloaded matching to the GPU: a CUDA brute-force search over coastline data, followed by dedup\n- Verified candidates with an open-rectangle constraint, coral-cay compactness checks, NDVI vegetation, and elevation/mountain checks\n\nTakeaway: when clever modeling stalls, just parallelize the brute-force search. All code and the final report are reproducible on GitHub.

Original post →

More from Fun

Fun channel →