Astronaut photo in. Coordinates out. No GPS.
Argus is a small satellite whose camera looks down at Earth. To work out its own orbit, it needs to know exactly what ground it is looking at: which pixels sit on which latitude and longitude. Those tie points go straight into orbit determination.
The old approach was closed-set: a classifier guesses a region, then landmark detectors hunt for features they were trained on. If it sees something new, it still guesses, and the wrong answer slides quietly into the orbit solution.
This project swaps that for an open-set, retrieve-then-match pipeline:
Every result on this site is a real output of that pipeline, run on the lab GPU and replayed here. The astronaut photos are harder than the real job: handheld, oblique, and rotated. Argus's own camera points much closer to straight down.
The retrieval model follows AstroLoc (Berton et al., 2025): paired + clustered multi-similarity losses, trained on NASA GAPE astronaut photos with machine-labeled footprints. Nothing is re-hosted here: each astronaut photo loads live from NASA's Gateway to Astronaut Photography of Earth (courtesy of the Earth Science and Remote Sensing Unit, NASA Johnson Space Center), and each reference tile is rebuilt from Sentinel-2 cloudless 2021 by EOX (CC BY-NC-SA 4.0), the same imagery the reference map was built from. This site only ships the pipeline's outputs. The gallery on this site is a seeded random sample that keeps every fix and some of the no-fixes. The headline numbers above come from the full unfiltered run.