1
$\begingroup$

I am working on a multi-class classification problem using images. I have a training library of images containing 9 different classes of object, however I will also need to train my image classifier to detect negative image examples (i.e. the image contains none of the 9 objects classes in the training library).

Assuming image frequencies in my training library are perfectly balanced, and each class contains n images, what is the best number of negative examples to include in the dataset? Intuition tells me to create n examples for the 'negative' class. Could anyone comment on the appropriateness of this?

Also if anyone could point towards any academic work published on this topic, I would be very grateful.

Thank you.

$\endgroup$
0
$\begingroup$

It is advisable to use equal proportion of instances if your n( number of examples) is a small number or I would say when n is not a good representation of the population. But this practice generally does not give good results when testing it on different set.So it would be better to look for the general trend among negative images and other images, what is their ratio is generally in reality because that might give you a hint of the optimal ratio which should take and if are not sure with what should be the optimal ratio, then you can choose any ratio of negative class and according to that you can assign "class_weight" parameter value while training the model.

$\endgroup$
  • $\begingroup$ Thanks for the input. Just to clarify I understand what you're proposing; the best solution 'in practice' would be to choose a number of negative examples such that the ratio of negative to positive examples in the real world is replicated in the training data set? $\endgroup$ – RDG Jul 28 '17 at 9:50
  • $\begingroup$ @RDG. Yes I meant the same $\endgroup$ – Harshit Mehta Jul 28 '17 at 14:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.