Image Annotation Services

Precision-labelled image datasets for computer vision teams — from fast bounding-box passes over millions of frames to pixel-accurate segmentation where the boundary is the whole point.

The label schema decides your model ceiling

Before any annotation begins, the most consequential decision on an image project is what the classes actually are. A schema that collapses two visually distinct objects into one class puts a hard ceiling on what the model can learn, and no amount of additional data lifts it. A schema that splits a class too finely produces categories with too few examples to train on and drives annotator disagreement.

We work through the schema with your team before production starts: what counts as an instance, how partially visible objects are handled, what the minimum size threshold is, and which cases are explicitly out of scope. That specification is written down, worked through against a pilot batch of your real images, and revised before it is applied at volume.

Annotation types we deliver

Different vision tasks need different geometry. Choosing the cheapest one that still supports the task is usually the right call — paying for pixel-accurate masks when an axis-aligned box would have trained the same detector is a common and expensive mistake.

Our image annotation work covers:

  • 2D bounding boxes — the standard for object detection. Fast, consistent, and sufficient wherever the task is "where is it and what is it".
  • Polygons and polylines — tighter than a box for irregular shapes, and the right choice for lane markings, road edges, and elongated objects where a box is mostly background.
  • Semantic segmentation — every pixel assigned a class. Used where the extent of a region matters: drivable surface, vegetation, sky, water.
  • Instance segmentation — pixel masks that also separate individual objects of the same class, for counting and per-object reasoning.
  • Keypoint and landmark annotation — skeletal joints, facial landmarks, and structural reference points for pose and alignment models.
  • Image classification and tagging — whole-image or multi-label tagging, including hierarchical taxonomies.

The edge cases that decide dataset quality

The difference between a usable dataset and an expensive one is almost entirely in how consistently the hard cases are handled. Occlusion is the clearest example: if half your annotators box the visible portion of a partly hidden object and half box the inferred full extent, the model receives contradictory supervision for exactly the situations it will face most often in deployment.

We resolve these explicitly in the specification rather than leaving them to annotator judgement — occlusion and truncation rules, minimum object size, how to treat reflections and screen-displayed objects, crowd and group handling, and what to do with genuinely ambiguous items. Ambiguous cases are flagged for review rather than guessed at, so they surface as a decision for your team instead of as silent noise in the training set.

Quality assurance

Every delivery passes through a layered review rather than a single pass. Annotators work to a written specification agreed before production starts; a second reviewer checks a defined proportion of each batch; and a final quality gate samples the batch independently against the same specification. Disagreements are not silently overwritten — they are resolved against the guideline, and where the guideline turns out to be ambiguous, the guideline is amended and the affected items are reworked.

That last step matters more than it sounds. Most systematic annotation error is not carelessness; it is a specification that failed to anticipate a real case. Treating every disagreement as a possible guideline defect is what keeps error rates from plateauing partway through a project.

For geometric tasks we measure agreement using intersection-over-union against a reviewed reference set, and track it per class rather than only in aggregate — a headline accuracy figure comfortably hides one class performing badly.

Delivery formats

Annotations are delivered in the format your pipeline already reads, so no conversion step sits between delivery and training. We commonly export COCO JSON, YOLO, Pascal VOC XML, and per-class mask images, and will match a custom schema where your training code expects one.

Deliveries include the annotation guideline version they were produced against. When a schema changes mid-programme, that record is what lets you tell which portion of the dataset needs reworking and which does not.

Data handling

Client data is encrypted in transit and at rest, access is limited to the specialists assigned to your project, and every member of the annotation workforce works under a signed confidentiality agreement. Where a dataset carries additional handling constraints, those are agreed in writing before any data is transferred.

Common questions

Which annotation type should I choose for object detection?

For standard object detection, 2D bounding boxes are almost always the right choice — they train detectors effectively at a fraction of the cost of segmentation. Move to polygons when your objects are irregular enough that a box is mostly background, and to segmentation only when the task genuinely depends on object extent rather than location.

How do you handle objects that are partly hidden?

Occlusion handling is fixed in the written specification before production begins, because inconsistency here is one of the largest sources of systematic error. We agree whether to annotate the visible portion only or the inferred full extent, apply that rule uniformly, and record it with the delivery.

Can you work to our existing annotation guidelines?

Yes. Where you already have a specification we annotate to it directly, and raise cases it does not cover rather than resolving them silently. Where you do not, we build one with your team against a pilot batch of your own images.

What export formats do you support?

COCO JSON, YOLO, Pascal VOC XML, and mask images are standard. Custom formats are supported where your training pipeline expects a specific schema.

Discuss your image annotation project

Tell us what you are building and what your data has to support. We will come back with a specification, a pilot scope, and a realistic timeline.

Start a conversation