Matt1up/chicago-grantpark-photogrammetry

Chicago / Grant Park — Aerial Photogrammetry + Terrestrial Laser Dataset

2,751 aerial photos (45.0 GB) plus the terrestrial laser scans (9.3 GB, 43 stations) over Grant Park and downtown Chicago, June 2020. CC BY 4.0.

⬇ Download

→ huggingface.co/datasets/Matt1up/chicago-grantpark-photogrammetry

Browse the Files tab and take what you want — no account needed. The 54 GB of imagery and laser scans lives there because GitHub won't host files that size; this repo holds docs and checksums.

Sample image · command-line options

---

🏆 Winner — RealityCapture #RCmonthlyChallenge, August 2020

This reconstruction and its companion tree 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 Chicago model appears in the reel under a created by: @Matt1up title card.

---

Grant Park, June 2020

Lollapalooza was cancelled that year. A friend was playing the virtual one and wanted the Chicago skyline behind him, so I flew the park and the surrounding downtown over two days.

The laser scanner went on the festival field itself — the two softball diamonds and the floodlit courts where Perry's Stage normally stands. That ground got the detail. The city around it was backdrop.

Downtown was empty. That part is not repeatable.

The field, empty. Perry's Stage is built over those diamonds and courts.

The reconstruction

Grant Park flythrough — click to watch on Vimeo

Creation reel — the capture and processing pipeline, click to watch on Vimeo

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

What's in the dataset

Images	2,751 JPEG · 45.03 GB
Sensor	Hasselblad L1D-20c — 1" 20 MP CMOS (DJI Mavic 2 Pro)
Resolution	5467 × 3582 (2,297) · 5464 × 3070 (400) · 5366 × 3575 (54)
Lens	10.3 mm — 28 mm full-frame equivalent, f/2.8
Geotagging	GPS lat/lon/altitude in EXIF — valid on all 2,751, none missing
Bounds	41.867317 – 41.874759 N · −87.624511 – −87.619413 W
Altitude	183 – 315 m above sea level
Captured	27–28 June 2020, 06:52 to 19:49
Laser	241 files · 9.30 GB · 43 stations — see below

Capture groups

Images are named by the flight that produced them, so you can take a subset without downloading everything.

group	images	captured	altitude a.s.l.
Grid_Down_1	485	27 Jun, 11:33–11:49	240–289 m
First_Flight_	400	27 Jun, 06:52–07:32	221–310 m
Grid_Down_2	372	27 Jun, 11:55–12:14	281–298 m
Park_Overhead_	354	28 Jun, 13:47–14:06	183–298 m
Grid_Down_4	282	27 Jun, 19:59–20:14	202–285 m
Buildings_3	258	28 Jun, 06:55–07:13	211–315 m
Sunday_Night_Buildings	216	28 Jun, 19:33–19:47	203–214 m
Buildings_2	154	28 Jun, 06:36–06:48	206–314 m
Buildings_1	81	28 Jun, 06:17–06:23	225–315 m
Buildings_4	54	28 Jun, 06:04–06:17	189–311 m
Grid_Down_3	54	27 Jun, 19:57–19:59	234–235 m
Park_Trees_NEW	24	28 Jun, 13:51–13:53	184–193 m
Sunday_Night_Street	17	28 Jun, 19:48–19:49	211–216 m
total	2,751

Group names are the flight names from capture. Times and altitudes above are read from EXIF, not estimated. Note First_Flight_ is the only 16:9 group (5464 × 3070); everything else is 3:2 apart from Buildings_4 at 5366 × 3575.

The capture rig

Aerial capture with a DJI Mavic 2 Pro (Hasselblad L1D-20c). Terrestrial laser with a FARO Focus S150, concentrated on the festival field rather than spread over the whole site.

Scale

The drone images are registered to the laser point cloud. The laser is the scale reference — GPS was not used for scale.

Accuracy is best inside the laser coverage, which is the festival field. Outside that the photogrammetry extrapolates from that anchor, and drift increases with distance from it.

If you re-align these images on their own, you get correct geometry at arbitrary scale. To get back to metric you need the laser cloud as a registration target, or ground control.

Download

→ huggingface.co/datasets/Matt1up/chicago-grantpark-photogrammetry

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

One file, straight from a browser or the shell:

curl -LO https://huggingface.co/datasets/Matt1up/chicago-grantpark-photogrammetry/resolve/main/images/Buildings_1-1.jpg

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

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

Take part of it with --include: 'sample/*' (~650 MB, look before committing to 55 GB), 'images/*', 'laser/*', or 'images/Grid_Down_1*' for one flight.

Everything, as a git repo (needs git-lfs — this is 55 GB):

git clone https://huggingface.co/datasets/Matt1up/chicago-grantpark-photogrammetry

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

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

./scripts/download.sh --sample              # ~650 MB, look before committing to 55 GB
./scripts/download.sh --full                # everything — images, laser scans, sample
./scripts/download.sh --images              # all 2,751 images
./scripts/download.sh --laser               # the 241 laser scan files
./scripts/download.sh --group Grid_Down_1   # one flight
./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.

The GPS in EXIF will place the reconstruction roughly on the map, but it is not the scale reference — see Scale above.

Gaussian splatting

No camera poses ship with this set, so run COLMAP or GLOMAP first — GLOMAP does global structure-from-motion and is far faster than COLMAP on 2,751 images.

Vanilla 3DGS is built for bounded scenes of a few hundred images and will run out of memory on a city block. The large-scale variants are what you want: CityGaussian, VastGaussian, Hierarchical 3DGS, Octree-GS, Scaffold-GS.

Downsample to ~1600 px wide before training. Nothing trains at 5467 px, and the originals are here so you can pick your own resolution.

Notes

• Three sensor crops appear in the set. 5467 × 3582 for most of it, 5464 × 3070 for the First_Flight_ group, and 5366 × 3575 for 54 frames. Same lens throughout; the solver should still be told to treat them as one camera.
• These are Lightroom exports, not raw. EXIF records processing in Lightroom Classic 9.3.
• The renders carry a "Capturing Reality" watermark. It was a RealityCapture Challenge entry on a promotional licence. Preview renders only, never the dataset images.
• Several passes cover the same ground at different altitudes and times of day. A curated subset aligns faster than all 2,751.

Laser scans

Registered laser point cloud — Grant Park tree line and ground plane.

241 files, 9.30 GB, 43 scan stations, from a FARO Focus S150. This is the scale reference the images are registered to.

tier	files
quarter_res	209
half_res	32

Two things to know before you download them:

They are in RealityScan's .lsp format. RealityScan imports them directly with importLaserScanFolder, and RealityScan is free. Nothing else reads .lsp. An E57 conversion (the open ASTM standard, readable by CloudCompare, Metashape, Autodesk, Blender) is planned — open an issue if you need it and I will prioritise it.

They are downsampled. half_res and quarter_res, not full scanner resolution.

hf download Matt1up/chicago-grantpark-photogrammetry --repo-type dataset --local-dir ./chicago --include 'laser/*'

manifest/laser.csv lists every file with its station, resolution tier and SHA-256.

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.

Chicago / Grant Park Aerial Photogrammetry Dataset — Matthew Guertin, 2020.
Licensed CC BY 4.0. https://github.com/Matt1Up/chicago-photogrammetry-dataset

See CITATION.cff for BibTeX and academic citation formats.

Related

• Tree photogrammetry dataset — 812 images of a single tree, with camera poses.
• 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.