Video Annotation Services

Where it helps, a model can propose pre-labels, for example boxes on a familiar class. A person still accepts, corrects or rejects every pre-label against the written guide, and novel classes start from a human hand. Automatic output that never meets a reviewer is not a VIPRO delivery.

What a track ID promises

Each object gets one track ID when it first appears, and that ID never moves to anything else. If a forklift leaves the scene and a different forklift drives in, the newcomer gets a new ID, even if it looks identical. If the same person walks behind a pillar and comes out again, they keep their original ID. The guideline writes down which of these cases count as the same object for your project and how annotators decide when they can't be sure.

IDs are unique within a clip, never reused after an object has been retired, and kept consistent across every layer: the box, the pose skeleton and any attributes all point to the same identity.

Crowded and fast-moving scenes

Dense scenes such as crowds, traffic queues, shelves, or players clustered around a ball are where tracking errors pile up. We split these clips into smaller segments for annotators, overlap the segment edges so identities can be handed over, and have a reviewer confirm every ID across each segment boundary before the clip is put back together.

Tracking crowded, fast play is covered in our sports video annotation guide.

When to interpolate

Interpolation works when motion between keyframes is smooth and predictable: a car cruising at steady speed, a person walking across a fixed camera's view, a parcel on a conveyor. The annotator places keyframes where motion changes, and the tool fills in the frames between.

When to label every frame

Fast direction changes, sudden scale changes, camera shake, deforming objects such as hands, and moments of contact or occlusion all need denser keyframes, often one on every frame. Egocentric and sports footage spend much of their time in this zone, so their guidelines set a much tighter keyframe spacing.

Correction passes

After interpolation, a correction pass scrubs through every generated frame and checks the box or polygon still fits the visible object. Wherever it drifts, the annotator adds a keyframe rather than nudging single frames. That way the fix carries through to the frames around it and the track stays smooth.

Choosing a sampling rate

Detection models trained on video often need only sampled frames, a few per second, because neighbouring frames are nearly identical. Tracking models need continuous sequences so identities can be learned across time. Action and fine-motion models usually need the full native rate around the events themselves.

A common pattern is mixed sampling: sparse frames across quiet stretches and full frame rate in windows around events of interest. We agree the scheme in the pilot and document it in the delivery so your training pipeline knows which frames carry labels.

Two-stage review

Annotators self-check each track on playback before submitting. An independent reviewer then audits a sample of every batch. They check identity continuity, box fit on interpolated frames, attribute changes and event boundaries against the written rubric.

Errors are logged by type: ID switch, fragmentation, missed object, extra object, geometry drift, wrong attribute and boundary outside tolerance. Batches that miss the agreed threshold go back for rework, and recurring edge cases are added to the guideline.