CornerNet-Lite
CornerNet-Lite copied to clipboard
princeton-vl
About "add downsampling lyaer before the hourglass module and remove one in hourglass modue"
As we can see, hourglass is a symmetric structure, how can we do "We reduce the maximum feature map resolution of the hourglass modules by adding one more downsampling layer before the hourglass modules, and remove one downsampling layer in each hourglass module."
I don't understand what these sentences do, could someone explain it to me? Thanks in advance. Could someone draw a picture to explain?
If you compare the code for cornernet and cornernet-squeeze, you will notice that in cornernet-squeeze, they added an extra residual block before the hourglass modules (pre variable, L71 in cornernet-squeeze), which downsamples the input further (stride=2). They also reduced the number of hourglass modules by 1 (hg_mods variable, L75 in cornernet-squeeze).
