I get that the temptation is to treat 'turtle_id' as a label and approach this as a classification problem but that feels a little weird given the context. Adding new turtles requires re-training, and limited pictures make this quite tricky (although it helps that there are quite a few for some turts making this easier than a 1-image-per-individual task).

I've just tested a (badly implemented) siamese network that tries to learn representations by predicting whether two images are of the same turtle. It doesn't do great (~0.03) but I get the feeling that this is at least a path worth exploring. Is anyone else doing some interesting contrastive learning? Any other different approaches working out? Or are we all just seeing who has the biggest ensemble of classification models :)