Introducing warnings: catch issues before they matter

We’ve added a new Warning System to Segments.ai that helps teams catch common annotation mistakes early, before they ship, get reviewed, or slow down your pipeline.

Warnings are customizable, easy to trigger, and tightly integrated into the labeling and review flow.

Reviewing annotations at scale is hard. Simple mistakes, like overlapping cuboids or inconsistent object sizes, can slip through even careful QA.

Previously, there was no clean way to catch these issues in the moment. Reviewers had to rely on visual checks or build manual scripts to spot problems.

Now, you can define custom rules per dataset, and Segments will flag issues directly in the interface, with just one click.

In any labeling or review session, you’ll now see a Warnings icon in the top bar. Click it to open the new sidebar.

From there, hit Check sample, and Segments will run your configured rules against the current scene. Any problems found are listed in the sidebar, with links to jump straight to the relevant object or frame.

You can:

• Dismiss individual warnings
• Re-run the check at any time
• Filter by sensor, frame, object, or warning type
• Group by type, sensor, or track

And when you save or submit a sample, Segments will prompt you if any warnings are still unresolved, so nothing slips through by accident.

ℹ️  This is currently available only in the 3D point cloud cuboid interface.

You can define rules for each dataset under Dataset Settings → Labeling → Warnings.

Two rule sets are available today:

Flags objects that overlap in 3D space, with options to exclude certain categories or category pairs (e.g., person + bicycle). Useful for ensuring clean geometry and avoiding impossible scenarios.

Warns if objects are too small, too large, or oddly shaped, with category-specific or global thresholds.

Configuration is quick and flexible. Only the fields you set are enforced.

→ See full guide on configuring rules

The warning system is live for all users. Start by configuring rules on your most important datasets, then open the sidebar and check your samples. You’ll be surprised at how much gets caught and how much faster the review process can be.

Let us know what rules you’d like to see next.