> For the complete documentation index, see [llms.txt](https://docs.unitlab.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.unitlab.ai/documentation/annotations/image-annotation.md).

# Image Annotation

Image annotation is a policy decision expressed through geometry. Choose the shape from the downstream task backward and write boundary rules that reviewers can apply consistently.

### What this editor is designed to preserve

* Source pixels, scale, and orientation.
* The ontology class, geometry, properties, relations, and required values.
* A reviewable boundary decision for ambiguous or partially visible objects.

![Image annotation Workbench in dark mode](https://292810646-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FGjVLUz4wthGkGlRKM6rM%2Fuploads%2FJR6FTWW8jqHz0VT6Hh7F%2Fimage-workbench.png?alt=media\&token=bc9b4805-7ab8-4b94-a6b8-4a0863a388d8)

*The image editor combines the canvas, ontology tools, item context, and workflow action in one task view.*

### Read the Workbench before annotating

The image workbench supports a broad geometry set:

* bounding box;
* cuboid;
* brush and eraser;
* polygon;
* mask-oriented Magic Touch;
* skeleton;
* line;
* keypoint;
* crosshair;
* comments.

Editing controls include pan, undo/redo, zoom/reset, image settings, shortcuts, selection, class reassignment, property/relation editing, deletion, and object stacking actions such as bring to front, bring forward, send backward, and send to back.

### Choose geometry from the model backward

* Use **bounding boxes** when coarse localization is sufficient and edge precision would not improve the model.
* Use **polygons or masks** when boundary shape, area, occlusion, or pixel-level separation matters.
* Use **cuboids** when a 3D-like image representation is required.
* Use **keypoints or skeletons** when the spatial arrangement of landmarks is the target.
* Use **lines/polylines** for elongated structures or paths.
* Use **item properties** for whole-image classification, captioning, or attributes that do not belong to one object.

In captioning workflows, the workbench presents item-level ontology properties instead of irrelevant spatial drawing tools. The ontology determines the work required for the item, so annotators see controls that match the labeling task.

Cuboid is available as a pseudo-3D eight-corner visual geometry in image, video, and document annotation, with shortcut **N**. UUEF preserves the cuboid geometry. COCO and YOLO do not have a native cuboid representation, so teams should choose UUEF when that geometry must remain lossless.

### Image quality is a policy problem

A technically valid polygon can still be wrong if annotators disagree about holes, shadows, truncation, reflections, or occluded boundaries. Production quality therefore requires:

* a geometry policy;
* hard examples in project instructions;
* an explicit treatment of ambiguity;
* reviewer calibration;
* release-level checks by class and source domain.

### From assigned task to review

{% stepper %}
{% step %}

#### 1. Read the task context

Confirm source identity, workflow stage, Instructions, and active ontology.
{% endstep %}

{% step %}

#### 2. Choose the class and geometry

Use the geometry required by the downstream use case, not simply the fastest tool.
{% endstep %}

{% step %}

#### 3. Create and refine the instance

Draw, zoom, pan, adjust vertices or mask boundaries, and handle occlusion according to policy.
{% endstep %}

{% step %}

#### 4. Complete structured values

Fill required properties, Item Properties, classifications, and relations.
{% endstep %}

{% step %}

#### 5. Inspect the full image

Look for missed instances, overlaps, duplicate identity, and systematic boundary drift.
{% endstep %}

{% step %}

#### 6. Save and route

Save the annotation state, then submit, reject, escalate, skip, or mark invalid only through the current stage action.
{% endstep %}
{% endstepper %}

### Modality-specific quality controls

| Decision        | Production guidance                                                                                 |
| --------------- | --------------------------------------------------------------------------------------------------- |
| Geometry choice | Match the downstream target: classification, detection, segmentation, keypoints, line, or relation. |
| Boundary policy | Define whether to include occluded, truncated, touching, or uncertain regions.                      |
| Object coverage | Scan the whole image at a useful zoom before completion.                                            |
| Required values | Resolve validation errors; do not invent values to force a valid state.                             |

### Route the result

Accepted work moves to the configured review or completion stage. Ambiguous policy routes to an issue or specialist; unusable data follows the invalid-data route; incorrect existing work follows rejection or rework.