Matt1up/drone-building-scans

Drone Building Scans

Two commercial buildings in the Minneapolis suburbs, flown August 2026. Full image sets with GPS intact, solved camera poses, and the finished 3D models.

Free drone photogrammetry test data for buildings barely exists — published datasets are either a whole city block or a turntable object, with nothing in between.

CC BY 4.0.

⬇ Download

→ huggingface.co/datasets/Matt1up/drone-building-scans

Browse the Files tab and take what you want — no account needed. The 16 GB of imagery lives there because GitHub won't host files that size; this repo holds docs, checksums and poses.

Textured model, 22 MB · sample image · command-line options

---

Building one — single storey commercial

Single storey. Brick and painted panel walls, glazed frontage, sidewalk, landscaping and striped parking. 951 images, all 951 align.

Building two — two storey commercial

Two storeys. Lap siding, exterior walkway with railings, external stairs, satellite dish. 395 images, 384 align.

What's in it

	flat-roof	shingle-roof
images	951	395
size	12.13 GB	4.52 GB
aligned	951 / 951	384 / 395
camera poses	included	included
tie points	2,230,994	1,025,000
reference model	.glb + .obj	.glb
scale	metric — AprilTag control points	unscaled

Capture

Both flown low and slow, under 16 m, gimbal sweeping from straight down to slightly upward. Nadir frames cover the top, oblique catch the edges, low and level frames get the walls. That is why the walls resolve — most aerial capture is nadir-only and building sides come out as smeared vertical texture.

Camera	DJI FC9313, 8.7 mm, f/1.8
Resolution	4096 × 3072 — native sensor readout, not an interpolated mode
ISO	100 throughout
Geotagging	GPS on every frame — 951/951 and 395/395
Filenames	original as written by the aircraft, EXIF untouched

Native resolution matters. These drones offer a higher-megapixel mode that interpolates from the same sensor well — it invents texture, and a photogrammetry solver treats invented texture as real observations. Everything here is the native readout.

Scale

The flat-roof scan is metrically scaled from a physical reference. A calibrated bar carrying two AprilTag 36h11 markers was laid in the grass and used as control points.

bar length	1.997 m, centre of tag to centre of tag
tag family	AprilTag 36h11
control points	poses/flat-roof/controlpoints.txt

The tags are visible in the imagery, so the scale is checkable from the data rather than something you take on trust — detect both tags, measure centre to centre, and it should come out at 1.997 m.

Watch the tag IDs. RealityScan labels 36h11 tags on its own scheme, not the official one. controlpoints.txt records them as 36h11:001 and 36h11:002 — those are RealityScan's labels. Under the official 36h11 numbering that most detectors (OpenCV, the apriltag library) report, the same two tags come back as 476 and 283. If you detect the tags yourself and get numbers that look nothing like the control point file, this is why.

The shingle-roof scan is not scaled. No control points, no physical reference in frame. Geometry is correct; absolute size is not established.

Camera poses

Both ship solved poses, so you can skip structure-from-motion:

<scan>/poses/xmp/        per-image XMP sidecars
<scan>/poses/colmap/     cameras.txt, images.txt, points3D.txt
<scan>/sfm/tiepoints.ply sparse cloud

Gaussian splatting

Poses and tie points are what 3DGS and NeRF pipelines ingest, so both scans train without running COLMAP first. Small enough that ordinary 3DGS handles them — no need for the large-scale variants a city block requires. Downsample to ~1600 px before training.

The reference .glb gives you something to compare a splat or a mesh against.

Reproducing

Both scans ship solved poses, so you only need to re-align if you want to compare. See docs/reproduce.md for settings and what to expect — notably that the shingle-roof alignment splits into three components and only the largest is published.

Download

→ huggingface.co/datasets/Matt1up/drone-building-scans

Click the Files tab and download whatever you want in a browser — no tooling, no account. The images and models live there; the GitHub repo holds the documentation, manifests, checksums and camera poses.

One file, straight from a browser or the shell:

curl -LO https://huggingface.co/datasets/Matt1up/drone-building-scans/resolve/main/flat-roof/model/flat-roof.glb

Everything, one command. Run it again if it stops — finished files are skipped.

pip install -U huggingface_hub
hf download Matt1up/drone-building-scans --repo-type dataset --local-dir ./drone-building-scans

Add --include 'shingle-roof/*' or --include 'flat-roof/*' to take one building.

Everything, as a git repo (needs git-lfs):

git clone https://huggingface.co/datasets/Matt1up/drone-building-scans

Or the helper scripts from the GitHub repo, which wrap the same command and add verify.sh to check every image against the published SHA-256 lists:

git clone https://github.com/Matt1Up/drone-building-scans-dataset && cd drone-building-scans-dataset

./scripts/download.sh --scan shingle-roof   # 4.52 GB, start here
./scripts/download.sh --scan flat-roof      # 12.13 GB
./scripts/download.sh --models              # just the reference models, ~500 MB
./scripts/verify.sh

More detail in docs/download.md.

Notes

• Both are commercial buildings. No residential property, no occupants.
• The licence covers the imagery. Trademarks or signage visible in it belong to their owners.
• Shot in evening light — long shadows across the parking areas in both sets.
• 11 of 395 images do not align on the shingle-roof scan. The flat-roof scan solves 951 of 951.

Licence

Creative Commons Attribution 4.0. Commercial use and ML training are fine. Credit required.

Drone Building Scans — Matthew Guertin, 2026. CC BY 4.0.
https://github.com/Matt1Up/drone-building-scans-dataset

Related

• Tree photogrammetry dataset — 812 images of one tree, with COLMAP poses.
• Chicago / Grant Park — 2,751 aerial images and 241 laser scan files over downtown Chicago.
• mattguertin.com