Regarding the UNet architecture using MobileNetV2 as the encoder

[YenYunn]

I would like to ask the author which layers of the encoder are used for skip connections with which layers of the UNet decoder.

Thank you.

[qubvel]

Hey @YenYunn, skip connections are taken right before each downsampling layer

[YenYunn]

Is its structure the same as the image?

[qubvel]

Similar, but not exactly, the number of channels in decoder is configurable and there is no last concat + two 3x3 conv layers, only 1x1

[YenYunn]

Okay, thank you.
I’d like to ask: Is the skip connection not directly connecting the input image to group 4 as shown in Figure ?
I took a look, and it seems there are only four skip connections in the bottleneck block. I couldn’t find the fifth one—may I ask where that one would be?

[qubvel]

Yes, there is no skip connection for the Group 4 block

[YenYunn]

So, we only have 4 skip connection ?
thanks