1
$\begingroup$

Faster RCNN consists of two modules.
(a) Region proposal network and (b) Fast R-CNN detector

The paper mention Region proposal network runs on the feature maps.
So, I have a input 256 x 256 image and I use VGG16 for example, after 5 times of pooling, the feature maps with only left size of 8 x 8.

My question is

  1. How am I going to propose on the feature maps if the size is not the same as the original image?

  2. Does the paper use any upsampling techniques to upsample the feature maps?

$\endgroup$
1
$\begingroup$

It is possible to propose on the feature maps even if the size is not the same as the original image. There is no upsampling needed for this. The RPN slides a small convnet over the 8x8 feature map, proposing k boxes at each pixel, where k is the number of aspect ratios.

When I say "propose" I mean that the network proposes a box by outputting 6 numbers: 2 of them are scores, and 4 of them are deltas, which we will use later.

The coordinates of each box can be computed as follows: Say at location 4, 5 on the feature map, the predicted delta of a box with size 64x128 was 10, -8, 9, 7

Then the upscaled corners of the box by default would be (4*32-64, 5*32-32) and (4*32+64, 5*32+32). The first term in each number comes from appropriately rescaling the box to the 256x256 image. The second term comes by halving the size of the box. Adding on the deltas we have (4*32-64+10, 5*32-32-8) and (4*32+64+9, 5*32+32+7)

$\endgroup$
  • $\begingroup$ Why is there 2 scores per anchor for object/no object instead of just one score? $\endgroup$ – Austin Jun 8 '18 at 11:52
  • 1
    $\begingroup$ @Jake From the paper: "For simplicity we implement the cls layer as a two-class softmax layer" $\endgroup$ – shimao Jun 8 '18 at 16:01

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.