Geolocating Random Islet Image Using Geometry & CUDA GPU Programming

On 16-08-2026, a report was written detailing a challenge called Gralhix 004, created by Sofia Santos. The challenge involved determining the name of a resort, the coordinates of an island, and the direction the camera was facing when a photo was taken. The author decided to solve the challenge using math and programming instead of using Google Lens, as they believed it was an opportunity to utilize their skills.

The photo in question was of a resort located on an island. Initially, the author attempted to obtain EXIF data from the image, but no useful information was found. The image depicted three landmasses, and the author struggled to create a correct perspective model of a bird's-eye view due to the drone's lack of elevation data. To estimate the relative distances between the three islands and angles of the triangle, the author built a small click GUI (01_triangle_gui.py) that recorded pixel coordinates for each point and calculated the triangle's geometry.

Next, the author used OpenStreetMap's land polygon set as a dataset to compare the image with real landmasses on Earth. After applying heuristic filters to remove landmasses outside the tropics, 141,131 land polygons survived the filter. The author then determined how many other centroids fell within 5km of a given point and capped the results at 10.

Each surviving point was then analyzed to find neighboring points within 20km, forming clusters of three or more. The maximum number of points per cluster was capped at 60 to avoid an exponential increase in combinations.

With 23,500 clusters producing 80,690,777 triples, each cluster had its triples processed by CUDA threads. Each thread sorted the three points by land area, identified the smallest point as P0 (the resort islet), and calculated angles and distance ratios using the cross product and dot product formulas. Only triples that met specific tolerance windows were considered valid.

To ensure the candidate islands formed a coral cay, additional filters were applied, including a Polsby Popper Score, minimum land fragments within a 1.5km radius, and a geometric filter on the polygon itself. After applying all filters, 948 candidates remained, with 158,784 passing the final mask test. The report concluded with a map of the 948 candidates, focusing on the resort islet and determining its shape resembled a coral cay.