About

Old imagery is only useful once you know where it is. Placing it on a map by hand means picking matching points one at a time, for every image. This tool finds those points automatically.

Status: experimental. This is an early version of the tool. It works well on some images and fails on others, and it says so when it is unsure. Treat results as a starting point, and check them.

Why it exists

It began as a desktop tool for placing declassified CORONA satellite photographs. Those scenes were taken by spy satellites between 1960 and 1972 and are now public. They cover places that have changed enormously since, which makes them valuable and makes them hard to align by eye.

The same problem comes up with archive aerial photos, scanned atlas sheets, field photographs and old satellite frames, so the tool is not limited to CORONA.

How it works

It uses two neural networks, SuperPoint to find distinctive points in an image and LightGlue to pair them across two images. A robust fit then turns the pairs into a transform, and rejects the bad pairs. Both networks run in your browser through WebGPU.

Running them in the browser is what keeps the cost low and your images private. Nothing is uploaded for processing.

What it is built on

• Copernicus Sentinel-2, © ESA, through Element 84's Earth Search over AWS Open Data.
• USGS NAIP aerial imagery of the United States.
• OpenStreetMap data, © OpenStreetMap contributors, ODbL.
• SuperPoint (DeTone, Malisiewicz, Rabinovich) and LightGlue (Lindenberger, Sarlin, Pollefeys).

Contact

Questions, bug reports and scenes that did not match are all welcome. Contact details will be added here before launch.