I am trying to understand a paragraph in the Faster RCNN paper.

We train and test both region proposal and object detection networks on single-scale images [7, 5]. We re-scale the images such that their shorter side is s = 600 pixels [5]. Multi-scale feature extraction may improve accuracy but does not exhibit a good speed-accuracy trade-off [5]. We also note that for ZF and VGG nets, the total stride on the last conv layer is 16 pixels on the re-scaled image, and thus is ∼10 pixels on a typical PASCAL image (∼500×375). Even such a large stride provides good results, though accuracy may be further improved with a smaller stride.

I would like to understand how they calculated the total stride on the last conv layer to be 16 pixels for the re-scaled image. I would like to decrease this stride to test accuracy improvements.

I'm also looking at the prototxt for training and testing and I can't figure out how he says the stride is 16 pixels.

Could you please show me how to calculate the stride being 16 pixels on the re-scaled image?


1 Answer 1


There are 4 max pooling layers in VGG before the last convolutional layer, each which has stride 2. $2^4 = 16$, so moving over by one in the feature map corresponds to 16 pixels in the input.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.