You could train a discriminator on top of your classifier, and feed your discriminator with both kinds of images (with and without objects). The discriminator would benefit from the classifier and you wouldn't need too many examples.
A simpler way would be to output probabilities per label instead of just labels, then threshold those with probability larger than a certain value, say, 0.9.
Or you could try to compute the entropy of the prediction vector, and if the entropy is too high, consider it "undecided".
A discriminator could be a great idea. Any idea on the kind of model efficient for such a task. I'm not sure a CNN is required for such a simple job.
Nowadays, I ouput label + probabilities, however detection threshold can sometimes be around 20% or 70%. I could of course design label-specific threshold, but it doesn't seem to be the correct way.
Or sometimes there is an object in front of me, but not any I know to classify. I would like to be able to detect that this object I don't know.
Concerning entropy, I would definitely give it a shot ! Very nice idea.
Any of you may know a system implementing such behaviour (e.g. I don't detect any object, I don't know this object) ?
A simpler way would be to output probabilities per label instead of just labels, then threshold those with probability larger than a certain value, say, 0.9.
Or you could try to compute the entropy of the prediction vector, and if the entropy is too high, consider it "undecided".