For classying images/objects CNNs are one possible or even the state-of-the-art solution but what if one wants to localize an object in an image?

I thought if I use only convolutional layers without (max)-pooling I should get feature maps which specialize on colors, edges and so on and if they would be trained for one specific object, they should be able to localize this object in any image?!

My problem now is figuring out what the cost function and teacher signal would be for such a network. I thought of e.g. two gaussian distributions, one for the x-coordinate and one for the y-coordinate of the object in the image. Would this be effective? Any better ideas?

Additionally I'm not sure how to model the teacher signal and cost function for such a scenario? Any tips?


Conv nets show substantial ability in localizing objects even without a specific training. This stands true particularly for CNNs endowed with global average pooling.

For more information, see:



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.