Label Quality Check¶

Analyze ▸ Label Quality Check…

Checks every labeled frame in the project and flags labels that are probably wrong. Running it before a training run is a much better use of ten seconds than almost anything else you could do with them — training on a systematically mislabeled set teaches the model to be systematically wrong.

Nothing runs when the dialog opens. It has two tabs, and each has its own run button:

Rules — click Run checks for the rule-based checks below.
Anomalies — click Run analysis for a statistical score per instance.

Click any result row to jump to that frame and instance. After you fix things, click Re-run to check again.

What it checks¶

Structural checks¶

Issue	What it means
Duplicate	Two instances on the same frame that are really the same animal — overlapping bounding boxes, or nodes sitting within a few pixels of each other
Incomplete frame	Fewer instances than the project's typical (median) count — an animal was probably missed
Negative frame	A frame marked as containing no animals that nevertheless has instances on it
Sparse instance	An instance with almost no visible nodes
Empty instance	An instance with no visible nodes at all
Out of range	Points outside the image bounds

Duplicate detection also catches the split duplicate case: one animal labeled as two instances with disjoint node sets — say head-and-torso in one instance, torso-and-tail in another. Their boxes may barely overlap, so plain IoU misses them; the check scores the combination of disjoint visible-node sets and body contact instead.

Geometric checks¶

These are the ones you cannot do by eye at scale.

Chirality (L/R) — a whole-instance left/right flip. This is invisible to distance-based checks, because a mirrored pose has all the same node-to-node distances. The check fits a body midline to each instance (the principal axis of its points) and asks which side of it each symmetric pair's "left" member sits on. It learns the canonical side from the majority of your instances, then flags the ones that disagree.

Declare your symmetries

This check needs to know which nodes are left/right pairs. Declare them as symmetries in the skeleton; failing that, the check infers pairs from node names that start or end with left/right or L/R, such as left_ear or ear_R. No pairs, no chirality check.

Chain order — nodes along a chain (a limb, a spine, a tail) labeled out of sequence. The check finds the longest chain in your skeleton and looks at the turning angles along it; a correctly ordered chain turns smoothly, a transposed one doubles back on itself.

Working through the results¶

On the Rules tab, results are a table of Issue, Frame, and Details. Click a row to jump to it and fix it.

Add flagged frames to Suggestions pushes every flagged frame into the Suggestions list, so you can work through them with Space like any other labeling pass.

Anomalies¶

The Anomalies tab scores every instance by how unusual its pose is compared with the rest of your labels — an unusual joint angle, edge length, or visibility pattern, a likely L/R swap, and the like. Click Run analysis; on a large project it shows progress and a Cancel button.

The table lists Score, GMM, Confidence, Top issue, and Frame, worst first. An instance is listed if either score flags it.
The Threshold slider sets the Score at which an instance is flagged. Lower it to flag more. The GMM column has its own fixed threshold (below).
Export CSV saves both scores for every instance, not just the flagged ones.

How the score works¶

The check learns what a normal pose looks like from your own labels, then measures how far each labeled instance is from normal. Predictions aren't scored.

For every instance it measures 18 things about the pose: edge lengths, joint angles, distances between nodes, overall size, how curved and compact the pose is, how symmetric the left/right pairs are, which nodes are visible, and how close it is to its most similar instance.
For each measurement, it works out how many standard deviations the instance is from the project's typical value.
The score comes from the most unusual of those measurements, scaled to between 0 and 1. A measurement 3 standard deviations out scores 0.5, and the default threshold of 0.7 flags instances about 4 standard deviations out.
Top issue names that most unusual measurement, such as Unusual joint angle or Likely L/R swap.

Confidence is read off the score: high above 0.8, medium above 0.5, low otherwise.

The GMM score¶

The Score looks at each measurement on its own, so it misses a pose whose measurements are each normal but whose combination is rare, such as one leg tucked in while the others are spread. The GMM column catches those.

It treats each instance's 18 measurements as one point and groups your labels into a few clusters of typical poses (up to 5).
It scores each instance by how far it falls from every cluster: the GMM score is the share of your labels that look more typical than this one.
Instances in the least typical 5% (GMM score 0.95 or more) are flagged, whatever the Threshold slider says. Their Top issue reads Improbable pose, followed by the measurement that stands out most.

The GMM needs at least 50 labeled instances; with fewer, the column shows a dash and only the Score is used. Because it always flags about 5% of your labels, treat its flags as poses worth a look, not as likely mistakes. The same labels always give the same GMM scores, but after you add or fix labels the borderline ones can change.

Reading the results

A flag isn't always a mistake. Unusual poses score high too, so look before you change anything.
Look for patterns. If one Top issue keeps coming up, the problem may be a labeling habit rather than single mistakes.
It works best with consistent labeling. The check compares your labels with each other.
Re-run it after you fix things, and after proofreading, to confirm the fixes.

When to run it¶

Before your first training run — cheapest possible model improvement
After a merge — merges are the main source of duplicates
After a bulk import — format conversions can put points out of range
Before publishing or sharing a dataset

Instance Size Distribution catches a different class of problem: instances whose size is anomalous, which usually means a stray point dragged far from the animal.