Datasets

All sources are packed into sharded tile stores at a fixed GSD (prepare_data.py, tools/pack_*.py). Each store records its GSD, so the loader knows the real pixel size of every tile.

Sources

Source	Sensor / type	GSD	Tiles (train / val / test)	Label	Role
GAMUS	US aerial, RGB (DC, NYC, PHL)	0.33 m nominal, 0.25 m measured	5004 / 859 / 2861	nDSM + 7 classes	main train, val and test
SynRS3D	synthetic cities	0.05–1.0 m in three families	12,536 train	nDSM + classes	pretraining and diversity
DFC23 Track 2	SuperView-1 satellite	0.5 m	1395 / 246	nDSM	out-of-domain val (v4)
India (labelled)	DFC23 New Delhi scenes	0.5 m	81 / 35	nDSM	near-domain val
US3D (DFC19)	WorldView-3 + LiDAR	0.32 m measured	2492 / 291, split by location	nDSM	v5
MVS3DM (IARPA)	WorldView-3 + LiDAR, San Fernando, Argentina	0.5 m, 512 px	train / val / test, split by location	nDSM (with trees)	v5, plus the LiDAR benchmark
NEON	aerial forest-canopy LiDAR	0.5 m, 1024 px	1846 / 118 / 232	canopy height (buildings masked)	v5 final: forests

SynRS3D is split into GSD families, each its own store: g005 (0.05–0.30 m), g05 (0.30–0.60 m) and g1 (0.60–1.00 m).

Download the data

Each source is available from its original provider. The packed tile stores used for training (already reprojected, tiled and normalised by prepare_data.py / tools/pack_*.py) are published on Kaggle, so a training run can start without re-running the preprocessing.

Dataset	Why it is used	Raw size	Original source	Packed for DepthWizard
GAMUS	Urban aerial RGB + heights; the main benchmark	64 GB	Hugging Face	Kaggle
SynRS3D	Synthetic scenes with exact heights; urban + terrain	23 GB	Hugging Face	Kaggle (g05 family)
DFC2019 Track 1 (US3D)	WorldView-3 + LiDAR; urban, sparse and trees	20.4 GB	IEEE DataPort	Kaggle
DFC2023 Track 2	Satellite imagery, including New Delhi scenes	12 GB	IEEE DataPort	Kaggle
MVS3DM (IARPA)	4-band WorldView-3 + LiDAR, a band layout close to Cartosat-2S; validation	22.6 GB	SpaceNet	Kaggle
NEON AOP	LiDAR canopy height; forested and hilly landscapes	~16 GB	NEON Data Portal	Kaggle

The trained model itself is on Kaggle Models. Check each provider’s licence before redistributing its data.

Training tiles per source (v4/v5 packing). MVS3DM is omitted because its per-split counts are not recorded in the repository.Source: Research-Paper/main.tex Table I; Model_Traning/v5/US3D_AND_GSD_NOTES.md; lightning/AFTER_NEON_UPLOAD.md

Label quirks that shaped training

• Trees at 0 m. DFC23 and the India scenes label about 97 % of green pixels below 1 m. In effect they teach “a tree is ground”. v5 first masked them with an excess-green index, and the final run dropped both sources.
• Stereo spikes. In DFC23, heights above 100 m are usually matching errors and can be marked invalid (max_valid_height_m).
• LiDAR spikes. MVS3DM and NEON rasters are despiked (|z − median₅| > 25 m → NoData) before packing.
• Coarse labels. NEON’s 1 m rasters on 0.5 m pixels are scored on 2 × 2 blocks (coarse-label term).
• Validation drift. v3 and v4 both printed val: 400 tiles from gamus, but from different packings, and v4’s set had 2.1× more tall pixels. v5 draws a seeded random 400 (seed 42) to make validation comparable (Design findings).

Measuring the real GSD

Height in metres depends on knowing metres per pixel, so the stored GSDs were checked against regulation-size sports fields visible in the imagery:

• tennis courts, 23.77 × 10.97 m
• American football fields, 109.73 × 48.77 m
• soccer penalty areas, 40.23 × 16.46 m

Sub-pixel line peaks were located on whiteness profiles. A measurement was kept only when an internal check (service line, yard lines, goal area) and the object’s two axes each agreed within 5 %.

Measured GSD per city (bars: median; grey ticks: interquartile range) against the nominal value stored with the data (wide tick). GAMUS measures about 0.25 m in every city, about 23 % finer than the 0.33 m it was stored with.Source: gsd_measurements/summary.csv

Dataset	City	Kept / tried	Median (m)	IQR (m)
US3D	JAX	14 / 126	0.3131	0.3098–0.3193
US3D	OMA	12 / 22	0.3374	0.3282–0.3458
GAMUS	PHL	9 / 10	0.2529	0.2527–0.2536
GAMUS	NYC	7 / 8	0.2507	0.2502–0.2510
GAMUS	DC	3 / 10	0.2562	0.2525–0.2602
US3D	ALL	26 / 148	0.3211	0.3129–0.3344
GAMUS	ALL	19 / 28	0.2527	0.2508–0.2535

Source: gsd_measurements/summary.csv

Impact on reported numbers

All GAMUS results before the v5 final run used the nominal 0.33 m. The v5 final run evaluates GAMUS at the corrected 0.25 m, so its GAMUS numbers are not comparable with earlier versions. Off-nadir US3D views can also be anisotropic: one view measured 0.344 × 0.378 m and was dropped.