Thanks for your suggestion.
PhotoStructure already has the best metadata inference and asset de-duplicating system that I'm aware of.
It scales to millions of asset files, works cross-platform, handles JPG, RAW, and transcodes videos to web-browsable formats automatically.
I haven't used an open-source project yet that didn't fall over after the first several thousand imports, and none de-dupe (not just the same SHA, mind you, but the semantically equivalent image, based on both metadata and rotation-invariant image content). They also tend to be abandoned when the original author moves on due personal reasons, toxic comments, or the "v2 rewrite" effort collapses under it's own weight.