@mfossati has been working on secondary models for logo detection, that classify an image into logo, signature, flag, coat_of_arms and some other classes. His original idea was to have primary and secondary classification, and class something as a logo if both primary and secondary models agree that it is

Another approach would be to add an out of domain class (basically to indicate that the image doesn't fall into any of the other classes), and to include training data for that when training the model. This ticket is to try that approach, and evaluate the results