Changepoint correction#

Once changepoints are detected, they can be used to refine label boundaries. The correction pipeline snaps hand-drawn label edges to nearby changepoints, producing more consistent annotations.

How it works#

The correction runs four steps in sequence:

Purge short intervals — remove labels shorter than the minimum duration.
Stitch gaps — merge adjacent same-label intervals separated by a small gap.
Snap boundaries — move each label’s start/end to the nearest changepoint, constrained by maximum expansion/shrink limits.
Purge short intervals (again) — snapping may create new short intervals.

Parameters#

Parameter	Description
Min label length	Labels shorter than this are removed.
Stitch gap	Maximum gap between same-label segments to merge.
Max expansion	How far a boundary can move outward toward a changepoint.
Max shrink	How far a boundary can move inward toward a changepoint.
Per-label thresholds	Override the global min label length for specific label classes.

These four parameters can be set in frames or seconds. I would recommend using frames for kinematics and seconds for audio data.

Automatic vs manual correction#

Automatic (during labelling):

Checkbox “Changepoint correction”: When enabled, label boundaries are snapped to changepoints as you draw them. This keeps hand-drawn annotations consistent in real time.

Manual (post-hoc, bulk application): Manual correction is especially useful for cleaning up model predictions in bulk — run the detector once, then snap all predicted boundaries to changepoints across the dataset in one step.

Single Trial: Applies the full correction pipeline to the current trial’s labels.
All Trials: Applies correction to every trial. The dataset is marked as corrected to prevent double-application.
Undo: Reverts the last correction (single or all trials).