# How to check a corrected video annotation delivery

Before you replace a video annotation delivery, compare the new version with the agreed correction. Keep the previous version available under your retention terms until you resolve unexpected changes. Check frame identity, annotation geometry and the clock your downstream task uses.

## Confirm which frames the supplier removed

Ask for the clip and source frame IDs that changed. If the export renumbers frames, ask for a map back to the source IDs. If media was cut, cropped or resized, agree how to map boxes into a common coordinate system before comparing them. A matching count cannot show which frames you received.

Ask whether each removed frame was excluded only from the delivery or deleted from the annotation task. In a five-frame fixture using CVAT v2.83.0’s interpolation method, either approach left frames 0, 1, 3 and 4. The four IDs matched, but the boxes on frames 1 and 3 shifted 15 pixels when frame 2’s keyframe was deleted before interpolation. Removing frame 2 only from the output preserved the original boxes. CVAT’s documentation advises reviewing tracks around deleted intervals. [Pinned interpolation method](https://github.com/cvat-ai/cvat/blob/678221a53665cded90e20822ba50704fd21e44ca/cvat/apps/dataset_manager/annotation.py#L1140-L1212) · [CVAT deletion guidance](https://docs.cvat.ai/docs/workspace/tasks-page/#how-to-delete-a-frame-from-a-task)

| Fictional correction agreement | Result on the retained frames | Decision on this check |
| --- | --- | --- |
| Omit frame 2 from delivery and preserve retained boxes | Frame 2 is absent and all retained boxes match | Pass the agreed-change check |
| Omit frame 2 from delivery and preserve retained boxes | Frame 2 is absent and boxes on frames 1 and 3 moved | Hold the revision and ask about the unexpected changes |

If the agreement allows changes to nearby retained boxes, review those boxes against the delivered frames. Matching the permitted change list only shows that the revision stayed within the agreed scope. The overlay check asks whether the new labels are correct. Keep a stable track ID or an explicit old-to-new track map so you can compare the same objects over time. [CVAT video annotation format](https://docs.cvat.ai/docs/dataset_management/formats/format-cvat/)

## Use the clock your task depends on

If the downstream task uses elapsed recording time, verify that the delivery carries source timestamps and identifies missing observations. If timing comes from a separate telemetry file, ask how those rows map to the delivered video frames. For a separate public CERES example, adjacent exported frames 267 and 268 were about 33 milliseconds apart on the 30 fps output grid. Their source timestamps were 41.2 seconds apart and neither endpoint carried the gap flag. The demo measures the stored timestamps, with physical sensor synchronization unmeasured. [Pinned CERES demonstration shard](https://huggingface.co/datasets/chrisvoncsefalvay/ceres-demos/tree/5c089ee6f91339e9cfa6d8b2ecbb267b03f93483/shards/episode-000001) · [CERES timestamp and gap fields](https://ceres.cam/documentation/data-description-and-output-format/)

Suppose a temporal model is allowed to use pairs of observations only when they are at most 100 milliseconds apart on the source clock. That illustrative rule would reject a pair across the CERES boundary. Computing the interval from the compact output timestamps would accept it. Ask for the source clock definition and state the maximum gap before creating time windows. If you cannot verify the required clock, keep the affected windows out of that time-dependent use until you resolve the evidence.

## Record the decision

| Check | Evidence to save |
| --- | --- |
| Version and correction | Previous and revised version, file inventory, agreed changes and unresolved discrepancies |
| Frame and track identity | Source clip and frame IDs, delivered IDs, track identifiers or a map, and any coordinate transform |
| Annotation geometry | Changed tracks and adjoining retained frames, expected coordinate changes, and the result of overlay checks |
| Timing | Clock and units, allowed elapsed interval, source gap markers and any windows excluded |
| Acceptance | Which checks passed, reviewer, open issues, and which version remains approved for use |

Apply the checks to the specific version you accept. A corrected set of examples can close the named issues while leaving the quality of the wider collection to a separate sample.

[Download the decision checker and saved results (ZIP)](https://sapien1.com/research/check-corrected-video-annotations-evidence/video-annotation-replay.zip)
