Does an image's background matter for detector/localisation in the training part (using CNN)?

For example, if I want to make a face detector, which one is better as training dataset?

  • Faces cropped dataset
  • Faces in a global scene dataset

Does it matter?


Of course, it matters.

Which one is best completely depends on your problem. The golden rule for machine learning problems is that you want the data you train on to be as representative as the data the model will encounter once live.

Typically, faces cropped dataset will be an easier problem because the CNN can directly focus on the face, but if your use case will use images of faces that are not cropped, your model might not perform well.

So use whatever data you expect to encounter later on. The only thing you could do is add a face-cropping algorithm before your network, and then you could train on the faces cropped dataset right away.

  • $\begingroup$ There is a sliding window and in methods using Hog and svm. And convutional part divide image in squarre which can be interpreted as a sliding windows no ? So cnn can find a face in an image even if it was trained with cropped face no ? $\endgroup$ – Jean Luc Jul 11 '18 at 9:54
  • 1
    $\begingroup$ It can, but you always run into the risk of having the background messing with your network, since this would be a type of input it has never seen before $\endgroup$ – Valentin Calomme Jul 11 '18 at 11:02

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.