How to set parameters"-source1.rfieldx" "-source1.rfieldy" and "-output.efieldx" "-output.efieldy" in otbcli_TensorflowModelServe

[chingrun]

Hello, I used otbtf to train my own model, and i got *.pb and the variable files, but when i use otbcli_TensorflowModelServe to infer, errors showed about "-source1.rfieldx" "-source1.rfieldy" and "-output.efieldx" "-output.efieldy". I am curious about how to set the parameter, in the paper i found the corresponding parameters were set to 80 and 16, how can i set the parameter in my own model? Thanks!

[remicres]

Hi,
These parameters depends on your deep net.
The "receptive field" (rfield) corresponds to the input volume that "sees" the deep net.
The "expression field" (efield) corresponds to the output volume that the deep net will create.

For instance:

• if your network performs single pixel classification (I called this "patch-based") e.g. it predict the class for the central pixel of a 16x16 patch, then your expression field is 1x1 (because your network generates 1 pixels for one input patch) and your receptive field is 16x16 (because your network inputs 16x16 pixels patches of input image).
• if your network performs semantic segmentation, you will likely have some expression field with a size > 1. For instance the fully convolutional network that you mention ("Maggiori" network in the paper) has a receptive field of 80x80 pixels and an expression field of 16x16.

You rfield and efield values (and spcscale if your net does change the physical spacing of the output) must be consistent with the model that you want to run.

I know I should put more documentation about the models provided in the repository. There is many examples to provide, to explain entirely how deep nets can be applied. I promise to add this very soon!

[remicres]

Just added some new documentation in doc/EXAMPLES.md

[chingrun]

Thank you for your kind help! @remicres