Custom models
Tuned on the project it will run on
We start from a finished model close to your data and finetune it on tiles out of the project it will run on, so it has seen your sensor, your density and your classes before it is handed over.
What we do to it
Your classes, tuned on your own data
Machine learning finds the classes and rule based steps settle what geometric context decides better: height thresholds for vegetation storeys, road rasters, your own vector files to separate one line from another.
Training runs in rounds against a test site you keep back, so accuracy is measured against ground truth the model has never seen. We reach over 98 % on many classes and projects. Once your catalog has been evaluated, over 95 % across it is the floor we commit to.
Sample areas from two parts of the project, around ten tiles each, with your class list. Every class has to be geometrically distinct and visible in every training set, so we take the density from the measured data.
Bring labels if you have them. If not, the platform is built for producing training data quickly, by your team or by ours.
Rounds against the held back test site, a short report after each. We adapt a classifier we already carry to your class list inside a single day. A new training takes a few weeks.
In the platform or straight through the API. The model stays yours to keep using and to retrain later.
On real data
A catalog that did not exist yet
Seawalls, groynes and the structures along a beach are in no standard catalog. Ten classes were defined with the council and trained on their own scans. Read the Worthing beach project.
Trained on the coast it runs on
Over 50 annotated areas and around 700 million points of training data, over about two months.
Deliverables
Where a trained model goes further
Every point in the cloud carries a class. That is what the filtering and the cleaning run on, and what every product under it is built from.
Your classes, to your specification. What is realistic on your data gets evaluated with you before anything is trained.
Height thresholds, road rasters and your own vector files separate what a model alone would blur. Those steps often work better than another model class.
Single 3D points where the geometry says they belong: pole tips, insulator suspension points, pole bases.
A custom model writes further dimensions beside the classification. Your original LAS or LAZ comes back with those dimensions written into it.
Photogrammetry, bathymetry, sonar and Geiger mode lidar. Each records the world in its own way, and a custom model is trained on the way yours does.
What lands in your hands
The output
- ALS
- Mobile Mapping
- TLS
- Photogrammetry
- Bathymetry
- Sonar
- Geiger mode lidar
at large scale
- LAS
- LAZ
- DXF
- Shapefile
- CSV
- GeoJSON
classified cloud and objects
Where it runs
On Microsoft Azure inside the EU by default.
A separate tenant in the Azure location your rules require.
The whole stack on your own hardware, air gapped where that is the requirement.
A project like yours
What the client said
“What used to be a cumbersome process, taking days of manual finetuning for the classifications can now be done in a fraction of the time - with higher consistency and greater accuracy.”
Questions
Before the classifier is trained
- How long does a custom model take?
We adapt an existing classifier to a new catalog in a single day, because the training base is modular. A new model takes a few weeks. Most of that time goes into generating the training data.
- What does it cost to run once it exists?
A model trained on your catalog runs at 0.20 € per million points. A standard classifier runs at 0.15 €. We quote the training per project, and the rate holds whatever your catalog contains.
- How much training data do you need from us?
It depends on how far your classes sit from the existing training base. We look at a sample of your data and name the amount before anything is agreed.
See it on your own data
Send a sample and we will run the classifier on it, so you can judge the result rather than the description.
