Field-tested, not generic
Our work is informed by robots operating on real Dutch farms — so annotators exercise judgement on the hard cases (overlapping canopies, disease, mixed rows) that generic crowd platforms get wrong.
Annotation · Sector
Annotation for agricultural robotics.
This is where we started, and it is still our deepest domain. We label the field imagery that teaches autonomous weeders, harvesters and crop-monitoring systems to tell a crop from a weed, a healthy plant from a diseased one, and exactly where to act — down to the millimetre.
Because we work alongside our parent company H2L Robotics BV and partner Cropr — who run autonomous laser weeders on real Dutch farms — our annotators see the same hard conditions the machines do: variable light, overlapping canopies, disease, and mixed crop rows. That field context is what separates a dataset that works in the polytunnel from one that works in a real field.
Annotation types
What the work looks like.
Plant & crop polygons
Pixel-accurate outlines around individual plants, following the real leaf boundary — the gold standard for segmentation and disease models where shape matters as much as location.
Weed vs. crop detection
Bounding boxes and instance labels separating crop plants from weed intruders in dense rows — the core signal an autonomous weeder is trained on.
Stem-base keypoints
A single point at each weed's base giving a laser weeder the exact coordinates to fire — millimetre precision, because a misplaced point kills a crop.
Growth-stage & disease
Labelling plants across the whole season, including disease-affected specimens, so detection stays reliable from emergence to harvest.
Coverage
What we annotate.
Crop detection
Weed detection
Plant polygons
Growth-stage
Disease / lesion
Stem-base keypoints
Row & terrain segmentation
Tulips · Potatoes · Chicory · Carrots · Onions
Where it's used
How it's used.
Where robots act
Stem-base keypoints and tight boxes give a laser weeder the exact coordinates to fire — sparing the crop, killing the weed. Precision here isn't cosmetic; it's the difference between a saved plant and a lost one.
Every growth stage
We label from emergence to harvest, including disease, so models stay accurate across the season instead of only on the imagery they were first trained on.
Agriculture remains our proving ground, and the QA process built here — multi-pass review, per-annotator quality tracking, consensus IoU — now underpins our work in every other sector.
