Classify all non recognizable objects into a single class.

I want to classify all objects that were not used during the training into a label as 'NonClassifiableObject'.

Following is an example that I will use describe what I want to express through my question.

Let us just say, my neural network has classification labels: Cat, Dog and Animal.

I want the network to classify cats as 'Cat', dogs as 'Dog' and any other animal that I MIGHT OR MIGHT NOT have used during training as 'Animal'.

All images of cats, dogs and any other animals both during training and testing will have same background and no other object apart from the animal will be in the image. There will no more than 1 animal in a single image.

I am at loss to how to approach this kind of training. The difficult part is to classify all objects that was clearly not used during training into a single label.

Thank you for your time.

[Rajesh Kontham]

I have come to realize that this is a more open-set recognition problem. I am still researching on this topic. Currently, I am trying to classify images as in-distribution, out-of-distribution and abnormal. The accuracy with which I can do varies upon the images used in training. So, I have much work ahead.

If anyone is working on these topics, I would like to have few discussions and sharing of ideas.

:) Thanks for your time.

[Gaia Cioffi]

Hi, I'm facing the same problem.

Have you found some solution or proper way to deal with this issue?