In the paper about Region Proposal Networks (RPN) (Faster R-CNN), Ren et al. explain the mini-network which classifies and regresses canditate region proposals (so-called) anchors, and visualize it with the following diagram:


What exactly does the intermediate layer look like? What does the "d" refer to in "256-d"? Pictured here, it looks like it is a fully-connected layer with 256 hidden nodes, but in this video it is shown to be a two-dimensional activation map. Which is it?

  • 2
    $\begingroup$ d is for dimension $\endgroup$ – shimao Jun 26 '19 at 9:36
  • $\begingroup$ @shimao Seems like that's sufficient for an answer! $\endgroup$ – mkt - Reinstate Monica Jun 26 '19 at 9:37
  • $\begingroup$ So is it a fully-connected layer with 256 hidden nodes? $\endgroup$ – EmielBoss Jun 26 '19 at 9:44
  • $\begingroup$ @shimao Hey man I'm really sorry, but it's unfortunately not sufficient for my limited brain. Could you clarify what kind of layer it is, exactly? $\endgroup$ – EmielBoss Jun 26 '19 at 20:04

Not sure if my right, Conv feature map for VGG is 14x14x512 (Layer 13 of VGG, right before the last max-pooling layer) so, we use 512 of 3x3 filter in the RPN (with 1 0padding), hence the output is still 14x14x512 --> 512-d

For ZFNet the cov feature map output is nxnx256 (i am not sure about the number of "n") thus we use 256 of 3x3 filter, hence the output is still nxnx256 --> 256-d

Being that said, i'm not really sure why author need to convolve using 3x3 filter again in the RPN, why don't just use the feature map and fed it into the cls layer and reg layer directly. Hope this answer part of your question. Thank You


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.