search

hszhao / PSPNet

Pyramid Scene Parsing Network, CVPR2017.
https://hszhao.github.io/projects/pspnet
Other
1.59k stars 545 forks source link

parameter settings for multi-scale testing #40

Open Vicky0522 opened 7 years ago

Vicky0522 commented 7 years ago

Hi, I'm confused with parameter settings when using multi-scale input the interpolation layer in PSP module is as follows: layer {
name: "conv5_3_pool1_interp"
type: "Interp"
bottom: "conv5_3_pool1_conv"
top: "conv5_3_pool1_interp"
interp_param {
height: 60
width: 60
}
} this can make scale of feature after pooling layers be the same as conv5_3 so they can be concatenated when input is 473 x 473 but how does it change when dealing with multi-scale input? I'm wondering about that, thank you if you can solve my question.

qizhuli commented 7 years ago

My understanding of the multi-scale testing code is that multi-scale testing is done by tiling 473x473 windows around an image if any of its two sides > 473, and then averaging all the outputs. So each forward pass still deals with just a 473x473 crop.