DataUp - dataset curation/insights + model evaluation on top of CVAT (show & tell)

[Ghallabi]

Hi CVAT community 👋

We’re building DataUp - an “intelligence layer” that connects to CVAT and adds dataset insights, versioned ground-truth snapshots, and model evaluation, without replacing CVAT.

What we offer today

• Connect CVAT Online → we automatically mirror your CVAT Tasks. In DataUp, each Task becomes a Dataset Card.
• For each Dataset Card, you can compute insights like:
  • blur + brightness buckets
  • Exact duplicate detection
  • text-to-image and image-to-image similarity search
  • Collections (save and reuse subsets of images for review, curation, or evaluation)
• Dataset versions: snapshots of ground truth (we don’t store images). Depending on whether you set up a read-only or read-write integration, you may also be able to roll back to previous ground-truth versions.
• Model evaluation:
  • register a model endpoint and evaluate it against a dataset version, or
  • run evaluation locally via our SDK and push results back to DataUp.

Why?

CVAT is excellent for annotation workflows, but many teams still rely on scripts to validate dataset readiness and run consistent evaluations. DataUp makes these steps repeatable and easy to compare across dataset versions. If you’d like to try it on a real CVAT task, we’re happy to help - and we also offer a free plan so you can get a feel for it.