Three annotation types, three failure modes
| Type | Typical failure | Control |
|---|---|---|
| Object tracks | ID switches after occlusion | Review every occlusion event, not random frames |
| Temporal segments | Inconsistent start/end boundaries | Written boundary rules with frame examples |
| Dense captions | Generic, repetitive language | Vocabulary guidance and spot-checks for specificity |
Interpolation needs its own review
Common video annotation tools support keyframe interpolation between manually set frames. Interpolation saves time but hides error between keyframes. A QA pass should sample interpolated frames specifically, weighted toward fast motion.
People in the frame
Large egocentric video collections such as Ego4D documented consent and de-identification procedures as part of the dataset release (Grauman et al., arXiv:2110.07058). Owners of archival footage should expect the same question: did the people recorded agree to this use?
Owner checklist for licensable video
- Recording consent or legal basis per collection site.
- Faces, plates and screens: blurred, retained, or consented?
- Annotation schema version recorded with each batch.
- Frame rate and resolution documented, including any re-encoding.
Have an annotated dataset or specialist expertise?
Tell us what you own, how the labels were reviewed and which rights you can license. A short overview of your dataset or annotation business is a useful place to start.
Email us about your dataset
Prepare for a licensing conversation Prefer to write directly? hello@callwavemedia.com