> 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

Unitlab’s image Workbench combines pixel-accurate geometry, reusable ontologies, interactive AI assistance, and workflow review for computer-vision training data. Choose the least complex geometry that preserves the signal your model needs, then make the decision reproducible through instructions and ontology rules.

{% hint style="info" %}
**Use this guide when:** you are building detection, segmentation, keypoint, pose, classification, captioning, or visual-relation datasets from still images.
{% endhint %}

## See image annotation in action

The current demo shows image labeling and AI-assisted proposal creation inside the same editable Workbench.

{% embed url="<https://homepage-files.s3.us-east-2.amazonaws.com/hero-videos/hero/auto-labeling-2.mp4>" %}

[Open the demo in a new tab](https://homepage-files.s3.us-east-2.amazonaws.com/hero-videos/hero/auto-labeling-2.mp4).

## Before you begin

1. Create or select a project whose data and ontology match this modality.
2. Confirm the project instructions define the unit of annotation, boundary or timing policy, required properties, and review route.
3. Open the project and enter the assigned item from the project data view or queue. The Workbench loads the modality-native editor inside the shared Unitlab shell.

See [Annotation Workbench](/documentation/annotations/annotation-workbench.md) for navigation, saving, item state, comments, issues, and workflow actions.

## Understand the image work surface

The image occupies the central canvas. The toolbar provides pan, Magic Touch, prompt-based detection, bounding box, cuboid, brush, eraser, polygon, skeleton, line, keypoint, crosshair, comments, image settings, undo/redo, zoom, and annotation shortcuts. The ontology panel exposes classes, class properties, relations, and whole-image Item Properties; the workflow action reflects the item’s current stage.

![Magic Touch selects a cherry and Find Similar proposes matching masks](/files/NPEIFwgND8aE98GUCcku)

*Interactive assistance starts from a human-selected object. Proposals stay editable and must be reviewed before acceptance.*

## Supported annotation model

| Annotation type                     | Use it for                                                                           |
| ----------------------------------- | ------------------------------------------------------------------------------------ |
| **Bounding box**                    | Object detection when approximate extent is sufficient.                              |
| **Polygon or mask**                 | Pixel-level boundaries, irregular shapes, area, or occlusion-sensitive segmentation. |
| **Brush and eraser**                | Fine mask correction, holes, thin structures, and local cleanup.                     |
| **Cuboid**                          | Perspective-aware pseudo-3D extent represented by eight visible corners.             |
| **Keypoint or skeleton**            | Landmarks and connected pose structures.                                             |
| **Line or polyline**                | Roads, contours, paths, and elongated structures.                                    |
| **Classification or Item Property** | Whole-image labels, captions, quality, scene, or acquisition attributes.             |
| **Relation**                        | A governed connection between annotated objects.                                     |

The ontology—not the file type alone—determines which tools and values the annotator sees. Use Item Properties for facts about the complete image and class properties for facts about a specific object.

## Pixel-accurate labeling

Zoom to the level required by the boundary policy, create a polygon or mask, and refine it with brush and eraser. Review thin structures, holes, touching instances, reflections, shadows, truncation, and occlusion consistently. Do not demand pixel precision when the downstream task only needs object localization; unnecessary detail increases review cost without improving the target signal.

![Pixel-level mask editing beside the full annotated image](/files/diFD52zLu1JfnHv3grqW)

*The full object and the detailed boundary remain part of one labeling decision.*

## Nested ontologies, properties, and relations

A reusable ontology can combine visual classes with required attributes, nested options, Item Properties, and relations. This keeps class meaning stable across annotators and projects. For example, a Vehicle object can require type and occlusion properties, relate to a Road object, while image-level weather and capture conditions remain Item Properties.

![Image ontology with classes, attributes, relations, and Item Properties](/files/MfKk41BsNzanR6kzwUmZ)

*Use ontology structure to make visual labels machine-readable and reviewable, not just visually correct.*

## AI-assisted image labeling

Magic Touch creates an editable segmentation proposal from an interactive prompt. Prompt Auto-Labeling can detect supported objects on the current image, and Find Similar uses a selected box, polygon, or mask to propose visually similar instances on that same image. Accept only after checking missed objects, false positives, boundary quality, class mapping, and overlap with existing annotations.

## Annotate one production item

{% stepper %}
{% step %}

#### 1. Orient to the item

Read the instructions, confirm the active ontology and workflow stage, then inspect the complete image before zooming into the first target.
{% endstep %}

{% step %}

#### 2. Choose the class and geometry

Select the ontology class or numeric hotkey, then activate the geometry required by the task. Do not substitute a box for a required mask or create object geometry for a whole-image Item Property.
{% endstep %}

{% step %}

#### 3. Create the annotation

Draw the box, polygon, mask, cuboid, line, keypoint, or skeleton. Zoom and pan while keeping enough surrounding context to interpret the object correctly.
{% endstep %}

{% step %}

#### 4. Refine and describe it

Adjust vertices or mask pixels, then complete every required class property. Add relations only between the intended source and target objects.
{% endstep %}

{% step %}

#### 5. Inspect the complete image

Scan for missed instances, duplicate objects, inconsistent class choice, boundary drift, and invalid overlap. Use visibility and object-order controls when dense annotations obscure one another.
{% endstep %}

{% step %}

#### 6. Save and route the item

Save the current state, then use the stage action to submit, approve, reject, escalate, skip, or mark invalid according to the project workflow.
{% endstep %}
{% endstepper %}

## Quality review

| Review focus           | What to check                                                                            |
| ---------------------- | ---------------------------------------------------------------------------------------- |
| **Coverage**           | Inspect the entire image at a useful zoom; do not review only the first dense region.    |
| **Geometry**           | Check that the chosen shape matches the downstream task and the project instruction.     |
| **Boundary policy**    | Apply the same rule to occluded, truncated, touching, reflective, and ambiguous objects. |
| **Ontology values**    | Resolve required properties, Item Properties, and relation validation.                   |
| **Assisted proposals** | Measure false positives, misses, and boundary errors before expanding automation.        |

{% hint style="warning" %}
A saved annotation is not automatically a production-ready annotation. Required values, boundary or timing policy, cross-item consistency, and the configured review stage still apply.
{% endhint %}

## Move from labels to governed data

Accepted image annotations move through the configured review route and into versioned dataset or release outputs. Preserve the geometry in an export format that supports it: cuboids, masks, nested properties, and relations may require a richer format than a detection-only export.

![Integrated Unitlab workflow connecting model assistance, annotation, review, and quality assurance](/files/yPoYNby78Kd9mxWTVqxy)

*Use workflows to keep model output, human correction, review, and approval in one traceable operating path.*

## Next steps

* Use [Detect Anything (SAM 1–SAM 3)](/documentation/auto-labeling/detect-anything-sam-1-sam-3.md) to calibrate interactive and batch assistance.
* Use [Multimodal overview](/documentation/multimodal-annotations/multimodal-overview.md) when related files or views must stay in one task.
* Curate difficult cases and review cohorts in [Data curation](/documentation/data/data-curation.md).
* Read the current [image annotation product overview](https://unitlab.ai/en/image-annotation) for the feature overview and current media.