GEOAI_3D

Geospatial-first AI workflows for 3D data. Lidar point clouds, photogrammetry, and Gaussian splatting, with coordinate reference systems, vertical datums, and accuracy reporting treated as first-class concerns rather than afterthoughts.

Status: pre-alpha (v0.1.0). The out-of-core georeferenced foundation is here: read and write LAS/LAZ, Parquet, and XYZ with coordinate-system and vertical-datum handling, geometric feature descriptors, and a streaming engine that computed features on a 722,964,406-point LiDAR tile — which would need ~21.5 GB to hold in memory — within 8.9 GB of peak memory on 4 CPU cores, bit-for-bit identical to the whole-cloud result. APIs may still change before 1.0. See the roadmap below and CONTRIBUTING.md if you want to be involved early.

Why this exists

Working with 3D geospatial data today means assembling a pipeline by hand from low-level libraries — PDAL or laspy for IO, Open3D for structures, COLMAP for Structure-from-Motion, gsplat for splatting — each excellent, none aware of the others, and none aware that your data has a coordinate system.

Three problems recur, and none of them are solved by any existing package:

Scale. A 200-million-point survey does not fit in memory. Chunked processing is straightforward until you need a geometric feature computed near a tile boundary to match the value it would have had on the unchunked cloud. GEOAI_3D treats that seam contract as a tested guarantee, not an implementation detail.

Vertical datums. Most 3D tooling reduces "CRS handling" to horizontal reprojection. Mixing GNSS ellipsoidal heights with an orthometric product silently introduces errors of tens of metres. GEOAI_3D refuses to guess: geoid separation is handled explicitly, and data without a declared vertical datum raises rather than defaults.

Accuracy. Geomatics is the discipline that quantifies uncertainty; most 3D AI tooling treats a coordinate as a fact. Per-point uncertainty, registration covariance, and propagation into derived products are intended to be properties of the data, not a separate analysis.

Design commitments

These are constraints on the project, not aspirations:

• Base install works on a CPU-only machine with no compiler. Anything requiring CUDA or a build toolchain lives in an optional extra, and CI verifies the bare install in a bare container.
• Cross-platform. Linux, macOS, and Windows are tested on every commit, across Python 3.10 to 3.13. Colab-friendly throughout.
• Spatial reference is never optional. Any function returning spatial data returns it with a CRS attached. Missing georeferencing is an error naming the parameter that would fix it, never a silent default.
• Build on the ecosystem, own the middle. PDAL, laspy, Open3D, COLMAP, and gsplat are dependencies, not things to reimplement. What GEOAI_3D provides is the out-of-core execution model, the datum and CRS layer, uncertainty propagation, and provenance that sit between them.
• Everything is provenanced. Outputs record what produced them, from which input, with which parameters and versions.

Installation

pip install geoai3d

Not yet published. Until the first release, install from source:

git clone https://github.com/agelamin17/geoai3d.git
cd geoai3d
pip install -e ".[dev]"

Roadmap

Development is staged. Each stage ships something installable and useful on its own.

Non-goals

Stating these early, because scope creep is the main risk to a project like this:

• Not competing with gsplat on rasterisation performance. GEOAI_3D wraps it and adds georeferencing.
• Not a desktop GUI application. Reach for non-coders comes via the QGIS plugin.
• Not an autonomous-driving toolkit. Aerial, terrestrial, and survey geometry are the target; automotive benchmarks and detectors are explicitly out of scope.
• Not a hosted service.

Contributing

Early contributors are genuinely welcome, particularly anyone with survey control data, benchmark scenes, or a workflow they would like to see supported. See CONTRIBUTING.md and SUPPORT.md.

Citing

A software paper is planned. Until then, cite the repository via CITATION.cff.

License

MIT. See LICENSE.

Code from proprietary or copyleft-incompatible sources is not accepted. See CONTRIBUTING.md for the specific traps.