Get startedData · Label and review
Use Annotation studio
Draw or correct object labels on images and sampled frames, then explicitly confirm their review.
Source reviewed 2026-09-16. Availability depends on your installation, permissions, and compatible runtimes. A supported path is not a guarantee of model quality or production readiness.
Outcome
Move from raw media to reviewed labels, linked regions, prompts, answers, and captions.
Start here
What you need to know first
Check the inputs below before starting. If you are new, begin with the first-project guide. A dataset holds media and labels; a model produces results; a deployment makes a selected model version callable. Creating one does not create the others.
Bring these inputs
- — A dataset version and review queue
- — Classes or a multimodal evidence schema
Expected output and limits
- — Geometry, classes, text spans, prompts, answers, or captions
- — Asset-level QA state
Core workflow
Open a dataset version
Open Build & evaluate → Annotation studio and choose the exact dataset version, or choose Review annotations on a dataset to open the currently selected image. Return to dataset restores that image and version. On phones, the image strip sits beneath the canvas; on desktop, standalone editing keeps it on the right. Standalone review uses Up/Down; embedded review uses Left/Right. Both save before moving and preserve the viewport. Draw a missing object and Score Studio will suggest its class, then preview visually matched boxes for explicit batch approval. Arrows inside text fields, menus, sliders, and dialogs keep their normal behavior.
Create evidence
Draw boxes, polygons, masks, or linked regions; add VLM text and answers where the schema requires them.
Accelerate carefully
Use Automatic first pass to discover dataset concepts and create draft boxes or masks in a protected new version. A drawn detection box uses a matching known object label when available; project-specific classes are inferred from a bounded sample of existing labeled crops. The detector runs once on the frame and only adds a focused crop pass when the object was missed; its session is reused and concurrent inference is bounded. Candidate boxes are normalized, pixel-sampled, and compared with the drawn object and sampled same-class boxes before they appear as similar suggestions. Ambiguous or low-similarity objects stay out of the batch; dismissing it keeps existing annotations unchanged.
Review and resolve
Saving keeps a draft. Confirm review or Confirm & next records approval. For an editable version with unreviewed images, Confirm all annotations appears below the image strip. Use it only after inspecting the whole set, including images with no objects; it confirms the version, not just visible thumbnails. Classes persist across the version.
What this surface supports
Boxes, polygons, masks, and class tools
Phrase grounding and VLM text
Automatic first pass across a complete dataset
Drawn-box class and similar-object suggestions
Viewport-safe keyboard review
Expert section
Contracts, signals, and failure modes
Use this section when you are defining acceptance criteria, automating the surface, or reviewing whether its output is safe to promote downstream.
Quality and operating signals
- — Confirmed versus unreviewed image counts
- — Saved draft state
- — Label coverage and visible geometry
Common failure modes
- — Accepting suggestions without review
- — Overlapping or malformed geometry
- — Text detached from its linked region
Expert release checklist
□ Inputs and dependencies are pinned to immutable versions.
□ Acceptance metrics include critical classes and operating slices.
□ Failure, retry, cost, and rollback behavior are understood.
□ The resulting artifact has an owner and a downstream review path.
Engineering safeguards
- — Suggested labels require review
- — Automatic work writes to an editable version, never the source
- — Undo and save state stay visible
- — Annotation and QA states are distinct