I am looking for a way to re-identify/classify/recognize x real life objects (x < 50) with a camera. Each object should be presented to the AI only once for learning and there's always only one of these objects in the query image. New objects should be addable to the list of "known" objects. The objects are not necessarily part of ImageNet nor do I have a training dataset with various instances of these objects.
In the beginning I have no "known" objects. Now I present a smartphone, a teddy bear and a pair of scissors to the system. It should learn to re-identify these three objects if presented in the future. The objects will be the exact same objects, i.e. not a different phone, but definitely in a different viewing angle, lighting etc.
My understanding is that I would have to place each object in an embedding space and do a simple nearest neighbor lookup in that space for the queries. Maybe just use a trained ResNet, cut off the classification and simply use the output vector for each object? Not sure what the best way would be.
Any advice or hint to the right direction would be highly appreciated.