GAMUS D4 validation

Full GAMUS validation comparison

DepthWizard v5 epoch-7 best.pt, the same checkpoint used in the landscape table, scored on all 859 GAMUS validation tiles. The D4 candidate averages eight rotations and reflections with 1.5× input scaling. Both settings use the same deterministic 512 px centre crop and the same valid pixels. The weights are unchanged.

Metric	Single pass	D4 + 1.5×	Change
Global RMSE	2.775 m	2.590 m	−0.184 m (−6.64%)
Global MAE	1.346 m	1.264 m	−0.082 m
Urban RMSE	3.211 m	3.061 m	−0.150 m
Sparse RMSE	2.210 m	1.253 m	−0.957 m
Forested RMSE	2.126 m	2.052 m	−0.075 m
Building RMSE	3.322 m	2.920 m	−0.403 m
Tree RMSE	4.164 m	4.058 m	−0.105 m
True heights ≥15 m RMSE	5.792 m	5.980 m	+0.187 m
30 m nDSM block RMSE	1.285 m	1.115 m	−0.170 m
Runtime	67.78 s	960.38 s	14.17×

Both candidates score the same 222,847,123 valid pixels. Zoom helps sparse tiles and buildings most. True ≥15 m heights worsen, and the earlier cross-dataset pilot showed US3D regression, so this setting is a GAMUS global-RMSE result rather than a universally safe default. The 30 m metric pools nDSM heights; it is not absolute DSM scoring. The held-out test split was not used.

The gain also appears outside the original 128-tile pilot: aggregate-derived RMSE on the other 731 validation tiles is 2.755 m single-pass versus 2.602 m with zoom. This remains validation evidence, not an independent test.

Evaluation details

Checkpoint: epoch 7 best.pt; 925 tensors loaded, none missing or unexpected.
Split: GAMUS validation, all 859 tiles; deterministic centre crop per tile.
GSD: 0.25 m source labels; 512 px model crop, zoom input 768 px.
Same preprocessing, BF16 and batch size 2; no model weights or target heights changed.
D4: four rotations, each with and without horizontal flip; inverse transforms averaged at the original output size.