Golden Gate Park. The trees are trees.
The model learned from lidar flown over Surrey, British Columbia, and was never trained on San Francisco. Here it labels as vegetation nearly every point the aerial photograph shows as vegetation.
more figures
In the park, over the seeds: to of the vegetation. The park has little the photograph shows as not vegetation, points, and the model labels of them vegetation; the method notes give why that figure is not read as a test.
A rule that calls every point 2 to 25 m up vegetation scores 100% on both sides, because the reference points are drawn from that height band. In the Sunset the model finds a little less of the vegetation than that rule, and takes in very little of the rest.
A point counts only inside a patch the photograph marks one way or the other, away from the patch's edge, because a tree in an aerial photograph leans away from the camera.
Where it learned
A Surrey tile held out of training: the model at home, where its labels came from.
more figures
Accuracy over the seeds: to .
Mean IoU (intersection over union: for each class, the points both the model and the label give it, over the points either gives it, averaged over classes): over the 15 classes ( to ), over the 6 coarse groups ( to ).
The same park on the old model
The old model read a fourth input, the return ratio. It was 0 on every point of DALES-2 as published on Hugging Face, and real on all but of San Francisco points. Same checkpoint, only the input changes, and the forest becomes buildings.
more figures
The photograph's not-vegetation points labelled vegetation: as delivered , with the return ratio at 0 .
Taller than anything it saw
The tallest 40 m block in training spans m of height. Here of blocks span more, up to m. Written down before training: buildings would score lower here than in the other built areas. They do, on of seeds.
more figures
Lift is the share of the null's shortfall from perfect precision that the model makes up: 0 is no better than the null, 1 is every building point right. Over the seeds: to here, and to elsewhere.
Precision against the footprints ( to ), the null's . Footprint covered ( to ), the null's .
The footprints scored here could not be registered to the lidar, so they take the offset measured in other areas.
Of the points more than 150 m up, labelled building ( to ), labelled pole ( to ).
Embarcadero. The bay is ground.
The training labels have no word for water or for a bridge, so the model says what it knows. The bay becomes ground, and part of the Bay Bridge deck becomes building.
more figures
Over the seeds: water labelled ground to , bridge deck labelled building to .