I am trying to use deep learning to classify the shape of properties/lots into categories such as: rectangular, square, irregular, corner, battle-axe (flagpole), and triangular.
I first combine road and lot/property data, then convert this to raster.
The raster has values as:
road = 1
property/lot = 2,
blank = 0 (I also used a -0.5m buffer to create small gap between lots).
I then created a new polygon layer and labelled each property to the correct shape, using values to represent the shape type; 1, 2, 3 ,4 etc. I then used the Export Training Data For Deep Learning, using the road and property as the raster file, and polygon as the input feature class.
This produced the attached 'example' image (exports to tiff, but provided here as pdf).
I trained the model at the same export tile size of 2560, and just used the single shot detector.
When I run the model, it works without issue (and creates vectors) however no useful result is achieved as show in the attached 'export'.
I am clearly doing something very wrong, but tutorials online seem to only be for satellite imagery and much smaller objects.
Any follow up questions, knowledge or advice would be appreciated greatly!
Thanks.