I have a CNN based object detector trained on WIDER Face dataset. It can successfully detect human faces in a given image.

Now, I am trying to detect abstract face and minimalistic face patterns in clouds or houses etc but having no success.

Initially, I thought I could lower the detection threshold to detect such patterns, but such scheme didn't work.

My question is there any way other than collecting and labelling such training examples (face like patterns) to solve this problem.


You probably need a training set that contains examples of abstract faces and minimalistic face patterns. There's no reason to expect a neural network trained on real faces to detect abstract face patterns; neural networks don't necessarily generalize that way.

| cite | improve this answer | |
  • $\begingroup$ I was hoping that they would have somehow learned general features. Now to see what features an object detector model has learnt, can we use the same techniques such as activation maximization or saliency maps as we use for classification models? $\endgroup$ – Asad Oct 17 '17 at 6:58
  • $\begingroup$ @Asad, I understand you were hoping that, but I don't think it's going to work (and as your experiment demonstrates, it doesn't seem to work that way). Visualizing what features a neural network has learned is a separate question that is best asked separately, after doing some research on your own. There are methods for visualization at the first layer, but unfortunately there's no really good solutions for that deeper layers. $\endgroup$ – D.W. Oct 18 '17 at 22:37

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.