about “ train_PQAN_pool2.prototxt"

[Hucley]

1.The output of pool_s layer isn't used in the next, why?
2. In this prototxt, you choose the "pool2/3x3_s2" as the input of quality generation unit, which is the different from the train_PQAN_image.prototxt. What's the difference between them?
##convolution
layer {
bottom: "pool2/3x3_s2"
top: "conv1_sss1"
name: "conv1_sss1"
type: "Convolution"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 32
kernel_size: 3
stride: 1
pad: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
bottom: "conv1_sss1"
top: "conv1_sss2"
name: "conv1_sss2"
type: "Convolution"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
convolution_param {
num_output: 32
kernel_size: 3
stride: 1
pad: 1
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "pool_s"
type: "Pooling"
bottom: "conv1_sss1"
top: "pool_s"
pooling_param {
pool: AVE

kernel_h:7
kernel_w:7
stride: 7

}
}

layer {
name: "fc1_s"
type: "InnerProduct"
bottom: "pool2/3x3_s2"
top: "fc1_s"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inner_product_param {
num_output: 3
weight_filler {
type: "gaussian"
std: 0.01
}
#weight_filler {
# type: "xavier"
#}
bias_filler {
type: "constant"
value: 0
}
}
}

[liuyuisanai]

@Hucley

1. pool_s is used for score generation.
2. train_PQAN_X.prototxt means feeding score generation unit with feature in level X. Please refer to Fig.8 in paper.

[Hucley]

”pool_s is used for score generation."
but I means .. in the prototxt, the input of “pool_s" layer is "conv1_sss1" and the input of "fc1_s" layer is “pool2/3x3_s2”. the output feature maps of the “pool_s" layer and "conv1_sss2" isn't used as the input in the next.

[liuyuisanai]

@Hucley
Sorry for the misunderstanding.
Branch *_sss is used for fcn-based score generation unit while fc1_s is used for fc-based score generation unit. In our experiments, we find that the latter one performs faster. However we still reserve the former one in prototxt for anyone who wanna to try the original quality generation unit as described in our paper.