I was inspired by the moon scan @nicolasdiolez made and wanted to try to replicate their results.
I used a custom data pipeline I made that embeds the images using CLIP to filter for moon images and clusters them together using EvoC, then filteres by exposure and sharpness, and deduplicates to get the final 328 images I used to train it. I didn't want to sift through ~11,000 images manually so I spent several days automating it to save a few hours of work lol. Though I am happy with the result!
I also made a high resolution traditional photogrammetry model in RealityScan that I may release at some point in the future.
cool detail: In the camera solve in RealityScan, you could clearly trace the arc of their orbital path in the solved 3D camera positions which was awesome.
COLMAP registration in RealistyScan.
Gaussian splat trained in LichtFeld Studio.
Original Photography by the Artemis II Crew.
Images courtesy of NASA Johnson Space Center.