About the generation process of flow field/displacement field

[DongChunL]

The term "flow/displacement field" mentioned below specifically refers to what is generated after the decoder, that is, it is between UNET and STN and not within the STN itself.

I see code like this: _self.flow=conv_fn(dec_nf[-1], dim, kernel_size=3, padding=1),_ but conv_fn is not visible. I understand that this sentence literally means that the flow field is obtained from the last layer of the decoder.Or maybe I was wrong?

However, I don't understand how this is done. What is the logic inside the conv_fn function? Can you provide the internal code of the conv_fn function? I haven't found it. Or, could you explain the conv_fn function in detail ,I mean, "
How to compute the flow field between images from feature maps?" Thank you.

[Pascal-bucketbreaker]

I have the inverse question: How to compute the terminal image from the original image and flowfield?

[adalca]

in which framework (tf/torch)

in tf, for example, we cover this in the tutorial at the end:

warp_model = vxm.networks.Transform(vol_shape, interp_method='nearest')
warped_seg = warp_model.predict([vol, warp])