Find Similar

Use one verified box, polygon, mask, or cuboid to find and review similar objects in the current image or frame.

Find Similar turns one verified object into a reviewed set of visually related proposals. It is designed for dense, repetitive scenes such as products, crops, cells, components, people, or vehicles.

Magic Touch selects one object and Find Similar proposes matching instances

What Find Similar does

Find Similar is a contextual Workbench action. It appears after you select a compatible object and searches the current image or current video frame. It does not search an entire dataset.

Seed geometry	Proposal geometry	Supported
Bounding box	Bounding box	Yes
Polygon	Polygon	Yes
Segmentation mask	Segmentation mask	Yes
Cuboid / 3D box	Cuboid / 3D box	Yes

The current experience exposes a confidence threshold, removes strong overlaps with committed annotations, and holds new results as pending proposals. Use Clear to discard the proposal set or Accept all after review.

Find Similar is example-driven. Prompt Labeling is text-driven. Both produce proposals, but they solve different discovery problems.

Before you start

• Choose a seed that is correctly classified and tightly annotated.

• Prefer a clear, representative instance rather than a heavily occluded edge case.

• Confirm repeated objects are visually similar enough for example-based retrieval.

• Zoom so the seed boundary can be inspected before search.

• Read the project policy for duplicates, partial objects, and minimum visible area.

Find repeated objects

1

Create or select the seed

Draw a box, polygon, mask, or cuboid around one representative instance. Select the finished object in the canvas or Objects panel.

2

Start Find Similar

Choose Find Similar from the contextual header action. If it is missing, verify the seed geometry is supported and the object is selected.

3

Tune confidence

Start with a conservative threshold. Lower it when recall is too low; raise it when unrelated candidates dominate.

4

Inspect pending proposals

Review the complete canvas, not only the area around the seed. Compare each proposal with the class definition and check overlap with existing annotations.

5

Accept or clear

Choose Accept all only when the set is appropriate. Otherwise clear the set, improve the seed or threshold, and run again.

6

Finish annotation QA

Correct boundaries, complete required properties, and submit through the configured workflow.

Current auto-labeling demo

Live Unitlab demo used on the Video Annotation product page. The result remains editable and reviewable in the Workbench.

Threshold strategy

Result pattern	Adjustment
Many false positives	Raise confidence or choose a more distinctive seed
Similar objects are missed	Lower confidence gradually and inspect the complete set
One object receives duplicate proposals	Confirm the seed is committed and inspect overlap suppression
Different states are mixed	Use a more specific class or split the work by state/property
Scale changes reduce recall	Seed a second representative scale and review it as a separate pass

Review controls

For each proposal, verify:

• class and instance identity;

• geometry tightness or boundary precision;

• truncation and occlusion policy;

• duplicates and overlap with existing objects;

• required class properties;

• relations to other objects;

• consistency with nearby manually labeled examples.

Similarity is not semantic proof. A visually close result can still violate the ontology, and a valid instance can be visually different from the seed.

When to use another tool

Need	Better choice
Find objects from a natural-language concept	Prompt Labeling
Detect all visible instances of the active class	Detect Anything
Propagate the same instance through frames or slices	Bidirectional Auto-Tracking
Search across many assets	Embeddings, similarity, and outliers
Process a project population before human review	Batch Auto-Labeling

Operational guidance

Find Similar does not consume the standard AI-inference quota in the current implementation. Treat that as an execution detail, not a reason to skip review. Measure accepted proposals and correction rate on a representative cohort before using it as a standard labeling step.