Is there a good way to understand how single-shot object detection works? The most basic way to do detection is use a sliding-window detector and look at the output of the NN to detect if a class is there or not.

I'm wondering if there is a way to understand how many of the single-shot detectors work? Internally is there some form of sliding window going on? Or is it basically the same detector learned at each point?

  • $\begingroup$ Single shot detectors are very black box, so you're not going to know how it works internally, all you can look at is the structure. I suggest looking at YOLO, and how it operates. You can find out more here: towardsdatascience.com/… $\endgroup$
    – Recessive
    Mar 3 '20 at 5:42
  • $\begingroup$ Most have internal regression layer which produce object box coordinates. $\endgroup$ Mar 3 '20 at 6:13
  • $\begingroup$ Most start with pretrained bounding boxes which are then regressed/removed/classified etc. as the layers go deeper. Most SSD's have different parts such as feature proposal and bounding box regressor, whereas yolo is a single network, SSD has multiple parts which you might have to look into. $\endgroup$
    – SajanGohil
    Mar 3 '20 at 6:57
  • $\begingroup$ As an obvious question, if the item of interest is moved some place else on the picture is the detection performance the same? It seems like a successful SSD would be very hard to train since position information is encoded in image which would result in poor generalization. $\endgroup$ Mar 7 '20 at 5:45

Your Answer

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

Browse other questions tagged or ask your own question.