DUC/STDN similar caffe implementation
Usage:
layer{
name: "enlarge"
type: "Enlarge"
bottom: "conv7_2" //size is 5*5
top: "conv6_2" //size is 10*10
enlarge_param {
size: 10 //size after enlarge
}
}
Note:
If the number of bottom channels entered does not divisible by c^2, I did a process similar to the mean pooling.Which should be different from STDN/DUC.