| --- |
| license: cc-by-4.0 |
| pretty_name: Single Tree High-Density Photogrammetry Dataset |
| task_categories: |
| - image-to-3d |
| - depth-estimation |
| tags: |
| - photogrammetry |
| - 3d-reconstruction |
| - structure-from-motion |
| - gaussian-splatting |
| - point-cloud |
| - drone |
| - uav |
| - aerial-imagery |
| - vegetation |
| - forestry |
| size_categories: |
| - n<1K |
| --- |
| |
| # Single Tree — High-Density Photogrammetry Dataset |
|
|
| **812 full-resolution photographs (14.1 GB) of one mature deciduous tree, captured from |
| ground level to above the canopy. 807 of 812 images align.** |
| Released under CC BY 4.0 — free for commercial, academic and ML use with attribution. |
|
|
| > ## 🏆 Winner — RealityCapture #RCmonthlyChallenge, August 2020 |
| > |
| > This reconstruction and its [companion Chicago city scan](https://github.com/Matt1Up/chicago-photogrammetry-dataset) 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 ▶](https://www.youtube.com/watch?v=PfzdaZbUrFc)**. |
| > |
| > 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. |
|
|
|  |
|
|
| --- |
|
|
| ## Why this exists |
|
|
| Vegetation is the hardest subject in photogrammetry. Thin branches, self-similar texture, |
| leaves that move between frames, and a canopy that occludes its own trunk — a tree breaks |
| assumptions that buildings never test. |
|
|
| Most published photogrammetry datasets are buildings, statues or turntable objects precisely |
| because those are easy. This one is deliberately the hard case: **a single tree, covered |
| densely enough to actually solve**, with the low-altitude trunk passes that most aerial |
| captures skip. |
|
|
| If you are benchmarking a matcher, a Gaussian splatting pipeline, or a mesh reconstructor, |
| this is the set that will tell you where it breaks. |
|
|
| ## The reconstruction |
|
|
| [](https://vimeo.com/485263810) |
|
|
| *Finished reconstruction — click to watch on Vimeo* |
|
|
| [](https://vimeo.com/494607575) |
|
|
| *Capture and processing — click to watch on Vimeo* |
|
|
| Full project write-up: **[mattguertin.com/portfolio/tree](https://mattguertin.com/portfolio/tree/)** |
|
|
| | | | | |
| |---|---|---| |
| |  |  |  | |
|
|
| ## What's in the dataset |
|
|
| | | | |
| |---|---| |
| | **Images** | 812 JPEG · 14.06 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 | |
|
|
| ### Capture tiers |
|
|
| The set is two sessions, two days apart, covering different heights. **The low and mid tiers |
| are the valuable, unusual part** — they are hover passes at knee and chest height with the |
| camera angled *upward*, capturing the trunk, root flare and canopy underside that a |
| conventional descending orbit never sees. |
|
|
| | group | images | date | height above takeoff | gimbal | covers | |
| |---|---:|---|---|---|---| |
| | `Original_low` | 85 | 2020-07-18 | +0.5 m | +6.2° (up) | trunk, root flare, canopy underside | |
| | `Original_mid` | 68 | 2020-07-18 | +1.7 m | +2.3° (up) | lower canopy, branch structure | |
| | `The_Tree` | 659 | 2020-07-20 | orbit to above canopy | varies | full crown and outer canopy | |
| | **total** | **812** | | | | | |
|
|
| ## The capture rig |
|
|
|  |
|
|
| Photography flown with a **DJI Mavic 2 Pro** (Hasselblad L1D-20c), developed from DNG in |
| Lightroom Classic 9.3 with consistent settings across the set. |
|
|
| A **FARO Focus S150** terrestrial laser scanner was also on site, visible above. |
| **Its data is not part of this release** — this dataset is the 812 photographs only. The |
| scanner is shown because it is part of the honest record of how the subject was captured, not |
| because point clouds are included. |
|
|
| ## Download |
|
|
| Images are hosted off GitHub — this repository holds documentation, manifests and checksums. |
| See **[docs/download.md](docs/download.md)** for mirrors and resumable download instructions. |
|
|
| ```bash |
| # sample pack first (~420 MB) — evaluate before committing to 14 GB |
| ./scripts/download.sh --sample |
| |
| # full image set |
| ./scripts/download.sh --full |
| |
| # just the low-altitude trunk tiers (153 images) |
| ./scripts/download.sh --group Original_low --group Original_mid |
| ``` |
|
|
| Every file is checksummed. After downloading: |
|
|
| ```bash |
| ./scripts/verify.sh |
| ``` |
|
|
| ## Reproducing the reconstruction |
|
|
| See **[docs/reproduce.md](docs/reproduce.md)** for alignment settings. |
| Aligns in RealityCapture / RealityScan, Agisoft Metashape, COLMAP and Meshroom. Images carry |
| GPS, so georeferencing works without ground control. |
|
|
| **Expect ~807/812.** A handful of frames genuinely do not solve — that is the honest result on |
| this subject, not a processing failure to debug away. |
|
|
| ## Known characteristics |
|
|
| Read these before you file a bug — they are properties of the capture, not defects in the upload. |
|
|
| - **Wind moved the subject.** Two sessions two days apart, outdoors, on a tree. Leaves and thin |
| branches are not in identical positions between frames. This is inherent to the subject and is |
| a large part of why the dataset is interesting. |
| - **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 lossless PNG |
| masters (~100 MB each, 15 GB total). They are not in this release because 8-bit is what every |
| photogrammetry pipeline actually consumes, and they would double the download for no |
| alignment benefit. Open an issue if you have a use for them. |
| - **These are Lightroom exports, not raw.** Raw DNGs are not part of this release. |
|
|
| ## Licence |
|
|
| [](https://creativecommons.org/licenses/by/4.0/) |
|
|
| Released under [Creative Commons Attribution 4.0 International](LICENSE). |
| **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](CITATION.cff) for BibTeX and academic citation formats. |
|
|
| ## Related |
|
|
| - **[Chicago / Grant Park dataset](https://github.com/Matt1Up/chicago-photogrammetry-dataset)** — |
| 2,751 aerial images and 43 laser stations over downtown Chicago. The large-area counterpart |
| to this controlled single-subject set. |
| - **[mattguertin.com](https://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. |
|
|
