Extract building footprints deep learning model returns awful results - what am I doing wrong?

Question

I have some high resolution (2cm) 8 bit unsigned imagery, and am trying to run Detect Objects Using Deep Learning using AGOLs USA Building Footprints dlpk, and the results are awful.  I'm curious if I'm doing something wrong.  My projection is StatePlane Tranverse Mercator.  See my screenshots below.

Jay_Gregory · Answer

So the issue with building extraction was a combination of the needing to extract the RGB (1 2 3) bands AND needing to downsample the imagery from 2cm to 30cm, which is what the model was trained on.  Once I did that, the output was much better (and I didn't need to save a new file).  Extracting the bands also helped improve the model output for pavement crack detection as well, though I didn't need to downsample for that model.

Jay_Gregory · Answer

This was an interesting suggestion: I ran extract bands to only extract 1,2, and 3, and reran the tool - the results were different but equally bad.  Don't know if I have to save it as new imagery (which I didn't).  I think it was running on the limited bands in-memory.

Anonymous User · Answer

...