In the early days of facial recognition, people would measure things manually. Can't you run your vector and algarithms against an instogram photo, but not scrape it as such, and just keep the vector data it's point to it? That's what gogle does as a search engine. It down loads the whole in er net, and then it makes an index of all common key words that you do that point to the original page. You don't actually have a copy of the photo. You have a pointer to the photo. If it gets removed, then you can't be helped by it. But you have this composite of it. Would that not be a way...
Check out Clearview AI: https://clearview.ai
FOLLOW Jason: https://linktr.ee/calacanis