There are many methods you can use to compare two images in ML (Siamese NN, CNNs, Ect.) What I cannot figure out is comparing a large number of images (Without Retraining) to find images of a different object. The best way I can describe this is a few shot learning problem without retraining.
My only real idea is to use an RNN and have it memorize some of the required features of an image while it parses through all of them. I would also likely have to ensemble multiple RNNs with different sets of images in case the first RNN starts off on the outliers.