I have 500-1000 labelled street level photos from which I need to train a model to guess the country. Most photos are of well known areas/landmarks, and I can assume the unseen photos will be similar. Nevertheless, unseen photos will probably have unseen areas/landmarks within them.
Apart from the standard approach of doing transfer learning on pre-trained ImageNet models, what might be some more problem-specific approaches I could take here?
Googleable terms or links to papers would be very much appreciated.