Geolocating an Island from an Image Using Geometry and CUDA Programming

Challenge Overview

The project aimed to identify a resort name, island coordinates, and camera direction from a single image of an island. Instead of using readily available tools like Google Lens, the developer opted for a solution based on mathematics and programming.

Initial Image Analysis

The first step involved checking the image metadata using `exiftool`, which revealed no useful information such as EXIF data, GPS coordinates, or camera model. The image was a WEBP file with dimensions 736x515 pixels. Without elevation data, the developer estimated relative distances and angles between the three visible landmasses: the main islet, a right island, and a left front island with a mountain peak.

Building a Geometric Fingerprint

A custom GUI tool, `01_triangle_gui.py`, was developed to record pixel coordinates of the landmasses and compute the triangle's geometry. To account for imprecision in manual clicking, a ±20% tolerance band was applied to the calculated values, creating a 'fingerprint' for the island configuration.

Searching Global Datasets

With the geometric fingerprint established, the next phase involved searching real-world landmass data. The developer used OpenStreetMap's `land-polygons-split-4326` dataset, which contains global coastline vectors in WGS84 format and is 882 MB in size. Heuristic filters were applied, starting with a tropical latitude bounding box (between -30° and 30° latitude) to reduce the dataset size, leaving 141,131 land polygons for further processing.