Paper mentioned in the question title deals with localization of certain objects in images. Paper mentions generating multiple types of object masks using deep neural network based on ImageNet, proposing multiple bounding boxes and scoring such proposals. However I can not find how the bounding box proposals are generated. If paper does not mention this method, what kind of algorithm would you suggest to generate such proposals?


Section "4 Detection as DNN Regression" explains that they used a neural network to obtain the bounding boxes:

enter image description here

  • $\begingroup$ This section mentions that they use DNN generate binary masks. I'm interested in the bounding boxes, called bb in the paper specified by their top left corner and bottom right corner. This is mentioned in section "5.2 Object Localization from DNN Output" but no method to obtain the bounding box propositions is mentioned, only their scoring. $\endgroup$ – Jakub Petriska Dec 12 '16 at 2:25
  • $\begingroup$ @JakubPetriska the binary mask IS the bounding box, you could calculate the box from this information, or maybe add a layer on top to do it for you. $\endgroup$ – Cristian Garcia Mar 29 '17 at 17:11
  • $\begingroup$ @CristianGarcia How would a layer generating the bounding box on top of the binary mask layer look like? $\endgroup$ – Jakub Petriska Apr 3 '17 at 9:24
  • $\begingroup$ @JakubPetriska I would propose a binary layer that learned to predict the position of e.g. the upper left corner and lower right pixels of the full (completely filled) bounding box binary mask suggested in the paper. In the end you are just doing very specific corner detection. $\endgroup$ – Cristian Garcia Apr 6 '17 at 17:11

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.