Urban 3D Challenge Dataset
This PR adds the Urban 3D Challenge dataset from the USSOCOM Urban 3D Competition (2017). This dataset combines 2D RGB imagery with 3D Digital Terrain Models (DTM) and Digital Surface Models (DSM)
Some notes on the dataset:
- 2048x2048 Maxar WorldView 0.5m RGB imagery
- DTM and DSM are both 2048x2048 single channel tiffs
- I added a precomputed normalized DSM (nDSM = DSM - DTM) added as this is generally what is used in the competition winner solutions
- Binary masks AND instance masks
- Plotting only binary masks for now as large number of buildings gives me OOM issues locally (open to suggestions)
Examples (left to right: RGB|DTM|DSM|nDSM|Target|Prediction):

I removed the trainer changes since it overcomplicates this PR. Will add in a future PR.
@isaaccorley we should finish this when you get a chance, I think it's pretty close to done
Looks like a bad rebase
Looks like a bad rebase
SOS
LGTM
Let me know when the remaining change requests have been addressed and I'll re-review.