Matt1up/tree-minnetonka-photogrammetry

Single Tree — High-Density Photogrammetry Dataset

812 photos of one mature deciduous tree, flown from 0.5 to 7.9 m above the ground. 15.1 GB. 807 align, and the solved camera poses ship with it in COLMAP format — point a Gaussian splatting pipeline straight at it, no structure-from-motion run required. CC BY 4.0.

⬇ Download

→ huggingface.co/datasets/Matt1up/tree-minnetonka-photogrammetry

Browse the Files tab and take what you want — no account needed. The 15 GB of imagery lives there because GitHub won't host files that size; this repo holds docs, checksums and poses.

Sample image · command-line options

---

🏆 Winner — RealityCapture #RCmonthlyChallenge, August 2020

This reconstruction and its companion Chicago city scan were both named winners of Capturing Reality's monthly challenge, announced by RealityScan — the makers of RealityCapture — on 17 September 2020 in Winners of AUGUST #RCmonthlyChallenge ▶.

The video description credits the win as "@Matt1up — Tree and Chicago city", and the tree model appears in the reel under a created by: @Matt1up title card.

---

The tree

There's a tree up the street from where I lived in Minnetonka. I watched the leaves out my window for a few days, and when they went completely still I'd grab the drone and the scanner, drive the three blocks up there, and fly.

A few times the wind picked back up and I packed it in and went home. I got two good sessions, 18 and 20 July 2020.

No wind is why it aligns.

812 photos, flown between 0.5 and 7.9 m above the ground. 807 align. Camera poses are included, so you can skip COLMAP.

The reconstruction

Finished reconstruction — click to watch on Vimeo

Capture and processing — click to watch on Vimeo

Full project write-up: mattguertin.com/portfolio/tree

What's in the dataset

Images	812 JPEG · 15.10 GB
Alignment	807 / 812 images solve
Sensor	Hasselblad L1D-20c — 1" 20 MP CMOS (DJI Mavic 2 Pro)
Resolution	5464 × 3640
Lens	10.3 mm — 28 mm full-frame equivalent, f/2.8
Geotagging	GPS latitude / longitude / altitude in EXIF, all 812 images
Location	Minnetonka, MN — 44.944 N, −93.426 W
Captured	18 and 20 July 2020
Camera poses	included — 807 solved cameras + 1.2M tie points (poses/)

Capture tiers

Two sessions, two days apart, at different heights. The low and mid tiers are hover passes at knee and chest height with the camera angled up, covering the trunk and the underside of the canopy.

group	images	date	height above takeoff	gimbal pitch
Original_low	85	2020-07-18	+0.5 m (fixed)	+4.2° to +8.2°
Original_mid	68	2020-07-18	+1.7 to +4.0 m	−8.0° to +2.7°
The_Tree	659	2020-07-20	+1.3 to +7.9 m	−53.1° to +17.5°
total	812

Heights and gimbal angles are read from EXIF across every frame in each group, not sampled. Positive gimbal is pointing upward.

The capture rig

Photography flown with a DJI Mavic 2 Pro (Hasselblad L1D-20c). EXIF records processing in Adobe Lightroom Classic 9.3.

A FARO Focus S150 was on site too. Its data is not in this release — this is the photographs only.

Camera poses

poses/ has 807 solved cameras with intrinsics and extrinsics, plus a 1,206,765-point sparse cloud. Same thing COLMAP would give you, so you can skip that step and go straight to Gaussian splatting (3DGS), NeRF, or meshing — poses plus tie points is exactly what those pipelines ingest.

poses/colmap/	COLMAP sparse reconstruction — cameras.txt, images.txt, points3D.txt
poses/xmp/	807 XMP sidecars, named to match images/
poses/cameras.csv	the same data as one table
poses/suspect_cameras.txt	4 cameras with implausible solved focal lengths
poses/unaligned.txt	the 5 that did not solve
tiepoints.ply	1.2M sparse points, 80 MB — hosted with the images, not in git

Solved in RealityScan 2.2, which is free, then converted to COLMAP format by scripts/xmp-to-colmap.py. The converter does not assume the camera-frame convention — it tests both by reprojecting tie points and keeps whichever puts them in front of the cameras. See poses/colmap/README.md.

hf download Matt1up/tree-minnetonka-photogrammetry --repo-type dataset --local-dir ./tree --include 'colmap/*'   # points3D.txt, 58 MB

The 5 that didn't align are all from the low and mid tiers, none from The_Tree. Four more solved to an impossible focal length and are listed in poses/suspect_cameras.txt — worth dropping before training.

Download

→ huggingface.co/datasets/Matt1up/tree-minnetonka-photogrammetry

Click the Files tab and download whatever you want in a browser — no tooling, no account. The images 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/tree-minnetonka-photogrammetry/resolve/main/images/The_Tree-1.jpg

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

pip install -U huggingface_hub
hf download Matt1up/tree-minnetonka-photogrammetry --repo-type dataset --local-dir ./tree

Take part of it with --include: 'sample/*' (~440 MB, look before committing to 16 GB), 'images/*', 'colmap/*', or 'images/Original_low*' for one capture group.

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

git clone https://huggingface.co/datasets/Matt1up/tree-minnetonka-photogrammetry

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 list:

git clone https://github.com/Matt1Up/tree-photogrammetry-dataset && cd tree-photogrammetry-dataset

./scripts/download.sh --sample     # ~440 MB, look before committing to 16 GB
./scripts/download.sh --full       # everything
./scripts/download.sh --colmap     # points3D.txt for splatting
./scripts/download.sh --group Original_low --group Original_mid
./scripts/verify.sh

More detail in docs/download.md.

Reproducing the reconstruction

See docs/reproduce.md for alignment settings. The images are ordinary geotagged JPEGs, so any structure-from-motion tool will read them — RealityScan, Metashape, COLMAP, Meshroom.

Expect ~807/812. The solved alignment is in poses/ if you want something to compare against.

Notes

• No wind. Both sessions were flown in still air. A windy recapture would not align the same.
• The 153 low/mid images had their metadata repaired. These were exported through RealityCapture, which stripped all EXIF. The original camera metadata — make, model, GPS, timestamp, exposure — was grafted back on from the untouched 16-bit source files. Pixel data is byte-identical to the export; only the metadata block was rewritten. Verified: decoded-RGB checksums match before and after.
• Filenames were normalised. Those same 153 files carried a RealityCapture double extension (Original_low-10.png.geometry.jpg). Renamed to Original_low-10.jpg. Content untouched.
• 16-bit originals exist for 153 images. The low and mid tiers have 16-bit PNG masters (~100 MB each, 15 GB total). Not included — they would double the download and no photogrammetry pipeline needs them. Open an issue if you want them.
• These are Lightroom exports. EXIF records Lightroom Classic 9.3. Original camera files are not part of this release.

Licence

Released under Creative Commons Attribution 4.0 International. You may use this commercially, and you may train models on it. You must give credit.

Single Tree Photogrammetry Dataset — Matthew Guertin, 2020.
Licensed CC BY 4.0. https://github.com/Matt1Up/tree-photogrammetry-dataset

See CITATION.cff for BibTeX and academic citation formats.

Related

• Chicago / Grant Park dataset — 2,751 aerial images and 43 laser stations over downtown Chicago.
• mattguertin.com — portfolio and other work.

---

Captured, processed and released by Matthew Guertin. If you build something with this, I would genuinely like to see it — open an issue.