Not exactly similar as the description in your paper

[marshallixp]

I have not found the 1/4 branch (low resolution) in your prototxt, but more important thing is that you used lots of 1x1 convolution instead of 3x3 to reduce the size with a bottleneck shape. Could you share more about your cascade guidance? Thanks a lot.

[howard-mahe]

Regarding the 1/4 low resolution image, you are right, it is not used in the provided prototxt. Instead there is a kind of hack of the 1/4 branch using the layer:

layer {
  name: "conv3_1_sub4"
  type: "Interp"
  bottom: "conv3_1"
  top: "conv3_1_sub4"
  interp_param {
    shrink_factor: 2
  }
}

Regarding the succession of 1x1 conv (reduce), 3x3x conv, 1x1 conv (increase), this is a classical ResNet module used for ResNet50+, as described by Kaiming He.

Finally, cascade guidance is just about deeply supervising the training process (see PSPNet article). Auxiliary loss at sub4 (res. 1/4) and sub2 (res. 1/2) permits to inject more gradients for low-level layers and thus improves the training.