How CVAT Keyframes and Interpolation Work in Video Annotation

The keyframe is the important decision

In a CVAT track, an annotator creates or corrects the object at selected frames. CVAT interpolates the geometry between those keyframes. A regular frame step can help an annotator navigate a sequence, but the frame step itself is not the annotation strategy.

A practical sequence might be: create the object when it first becomes valid, move forward until its position, scale, shape or visibility changes meaningfully, correct the annotation to create another keyframe, and continue until the object leaves the annotation scope.

When to add another keyframe

Add or correct a keyframe when interpolation would otherwise stop matching the visible object. Common triggers include:

• A change in direction or speed
• A significant scale or perspective change
• Partial or full occlusion
• A change in visible shape
• A camera cut, shake or rapid movement
• The object entering or leaving the valid annotation region

Predictable motion may require fewer keyframes. Irregular motion, deformation and crowded interactions require closer review and more corrections.

Track IDs must represent identity

Interpolation is useful only if the track belongs to the same physical object. When two similar vehicles cross or an object disappears behind an obstruction, geometry may appear plausible while the identity becomes wrong. Reviewers should therefore check the sequence, not only isolated frames.

The project guideline should state whether an object keeps its Track ID after temporary occlusion or receives a new ID after leaving and re-entering the scene. Without that rule, different annotators can produce internally consistent but mutually incompatible results.

Occlusion is not the same as absence

An occluded object may remain part of a track even when its visible geometry is reduced. An object outside the frame or outside the defined annotation zone may need a different state. Use the platform’s outside, occluded and keyframe controls according to the client’s specification rather than treating them as interchangeable.

Review interpolation between keyframes

Checking only the frames an annotator touched can miss errors between them. Sample the interpolated interval, especially around direction changes, occlusions, scale transitions and crowded interactions. For polygons, also check whether point movement causes edges to cross or drift away from the object boundary.

A practical QA checklist

1. Confirm the track starts and ends according to the guideline.
2. Check that the class and attributes remain correct.
3. Review Track-ID continuity through occlusion and crossing objects.
4. Inspect frames between keyframes for drift.
5. Confirm outside and occluded states are applied correctly.
6. Look for track gaps, duplicates and identity switches.
7. Validate the exported sequence and frame references.

Interpolation saves drawing, not judgment

The best use of interpolation is to automate predictable movement while keeping human attention on motion changes, identity, visibility and difficult frames. That balance produces faster annotation without treating every generated frame as automatically correct.

For projects that need persistent tracks, temporal events or an independent sequence review, see our video object-tracking service, temporal annotation service and dataset QA workflow.