Impossible G3D architecture

[robinchm]

Paper's G3D architecture seems to be plainly wrong.

Let's say feature map at the red circle has the shape (B, C, H, W). And ResBlock3D does not change its size. At the blue circle we are adding (B, C, H, W) to (B, C, 2H, 2W). Don't know where the issue is.

[johndpope]

ok - in progress.

[robinchm]

I end up moving the the shortcut in between the resblock3d and downsample. I think this makes more sense and the author just didn't bother to get such detail correctly present.

[flyingshan]

I end up moving the the shortcut in between the resblock3d and downsample. I think this makes more sense and the author just didn't bother to get such detail correctly present.

Hello, may I ask do you have any experimental results about this, or try other structures?

[robinchm]

I end up moving the the shortcut in between the resblock3d and downsample. I think this makes more sense and the author just didn't bother to get such detail correctly present.

Hello, may I ask do you have any experimental results about this, or try other structures?

I haven't even figured out the exact implementation of the network, of course no training was attempted.