CloudLabeller streamlines object annotation in 3D point clouds. It takes you from a folder of photos to a dense, fully labelled point cloud in one desktop app: reconstruct the scene photogrammetrically, label a handful of images — or paint the cloud directly — and let camera projection and a U-Net propagate the labels to everything else.
Manually classifying millions of points is the bottleneck of every photogrammetric mapping workflow. CloudLabeller attacks it from both sides: labels drawn on 2D images are projected onto the 3D cloud (visibility-aware voting across views), labels on the cloud are projected back onto every image, and a U-Net trained on the few images you touched predicts the rest. Built for geological outcrop mapping, but suited to any survey where per-point classes matter.
Photogrammetry (COLMAP / pycolmap)
- Structure-from-Motion with spatial (GPS), sequential or exhaustive matching; GPU SIFT + matching via CUDA COLMAP, with automatic CPU fallback.
- Dense multi-view stereo with resolution presets, a fast Draft quality mode, and crash-resumable workspaces.
- Meshing: Poisson (depth auto-matched to the cloud's density, with run-time and RAM estimates) or Delaunay, rebuilt automatically after MVS.
- Georeferencing to the images' EXIF GPS (metric, true-north local frame) — automatically after SfM (optional) or on demand, with outlier-robust fitting; runs natively, no COLMAP executable needed.
- Reproject the whole project (clouds, mesh, cameras) into any projected CRS — coordinates are stored minus a km-rounded offset to preserve precision, the current frame shows in the status bar, and camera positions appear next to the file names in the Dataset pane.
- A one-click full pipeline (SfM → MVS → mesh) for overnight runs — the machine is kept awake and every stage logs to the project.
- COLMAP is not bundled: Photogrammetry → Download COLMAP… fetches the official release on demand (CUDA or CPU-only, auto-detected).
Labelling
- Draw polygon labels on images or select points/faces on the cloud and mesh with a shared class schema (colours, hotkeys) and full undo/redo.
- Bidirectional transfer: Images → Cloud (multi-view voting) and Cloud → Images (per-pixel masks), both visibility-aware.
- An image-overlap (covisibility) graph from the SfM solve keeps every transfer restricted to the images that actually see the labelled area.
- Per-image status dots show label provenance at a glance: user-drawn, projected/predicted, or unlabelled.
Machine learning
- Train a U-Net on your labelled images (TensorFlow), predict the whole image store, and send the predictions to the cloud — label a few images, annotate the entire survey.
Data interchange
- Export labelled clouds to
.las,.plyor.csvand meshes to.ply,.objor.stlfor GIS and geomodelling tools. - Georeferenced projects export in any projected CRS — searchable EPSG
picker with the site's UTM zone pre-selected, optional sea-level (EGM96)
heights. The CRS ships inside the LAS header and as a
.prjsidecar for the other formats. - Projects are plain folders — copy them between machines as-is.
CloudLabeller is not published on PyPI — install it from this repository.
Run from source (Python 3.10; developed and tested on Windows):
git clone https://github.com/italo-goncalves/cloudlabeller.git
cd cloudlabeller
pip install -r requirements.txt # dependencies only, pinned — see notes inside
python -m cloudlabellerBuild the portable Windows distribution — a self-contained folder with its own embedded Python that runs on machines with no Python installed:
python scripts/build_portable.py --out C:/Temp/cl_buildThe script downloads the embeddable Python 3.10, installs the requirements
into it (~2 GB — be patient), copies the app and writes the launchers
(CloudLabeller.vbs / .bat); zip the resulting CloudLabeller folder to
distribute it. By default your local COLMAP bundle (~/.cloudlabeller/colmap,
if present) is included; pass --skip-colmap for a much smaller archive —
recipients can then fetch COLMAP in-app via Photogrammetry → Download
COLMAP….
An NVIDIA GPU is required for dense MVS (CUDA); everything else — SfM, labelling, training, prediction — runs on the CPU.
Development: architecture in DESIGN.md; tests with
QT_QPA_PLATFORM=offscreen PYVISTA_OFF_SCREEN=true pytest tests/ -q.
- File → New Project…, then Add Images… to fill the image store.
- Photogrammetry → Run Full Pipeline… (or run SfM / MVS / mesh individually). Georeference when the reconstruction is final.
- Label a few well-chosen images (or paint the cloud) and run Transfer → Images → Cloud.
- Model → Train U-Net…, then Predict All Images… and transfer again — refine and repeat until the cloud is clean.
- File → Export Point Cloud… / Export Mesh…
CloudLabeller is an AI-assisted project: most of its application code was written, tested and debugged by an AI coding assistant (Anthropic's Claude) under the direction of the author, who defined the requirements, the scientific approach and the machine-learning methodology. As with any software, validate results independently before scientific or operational use.
CloudLabeller is dual-licensed:
- GNU GPL v3 or later — see LICENSE. Free to use, modify and redistribute; software that incorporates or redistributes it must remain under the GPL.
- Commercial license — for use in closed-source products or redistribution outside the GPL's terms, a separate commercial license is available from the author: contact Ítalo Gomes Gonçalves on LinkedIn.
Third-party components keep their own licenses; binary distributions include
the full collection in THIRD_PARTY_LICENSES.txt. COLMAP is not shipped
with CloudLabeller — the app downloads the official release on demand
(Photogrammetry → Download COLMAP…).
To keep the dual-licensing model possible, external contributions can only be accepted with a copyright assignment (CLA) — please contact the author (via LinkedIn above) before opening a pull request.
If you use CloudLabeller in research, please cite the software — see CITATION.cff (GitHub's Cite this repository button generates BibTeX/APA from it).