Labeling Instances¶

An instance is one animal's pose on one frame: a position (and a visible/not-visible flag) for each node in the project's skeleton.

Adding an instance¶

How	What it does
`Cmd`+`I` / `Ctrl`+`I`	Add an instance using the current placement method
Labels ▸ Add Instance	Same, from the menu
Right-click ▸ Add Instance	Pick the placement method for this one instance
`Ctrl`-drag a node of an existing user instance	Clone it, pose and all (`Ctrl` on macOS too, not `Cmd`)

Placement methods¶

Labels ▸ Instance Placement Method decides the starting pose of a new instance. Getting this right saves a lot of dragging.

Method	Starting pose
Best	Centered template, offset to avoid overlapping existing instances (default)
Template	The skeleton's stored template pose, unmodified
Force Directed	Centered, then iteratively pushed away from existing instances
Random	Random node positions
Copy Prior Frame	The pose from the previous labeled frame
Copy Predictions	The predicted instance already on this frame

Once you have a model, Copy Predictions turns labeling into correction.

Editing¶

Drag a node to move it.
Move the whole pose: double-click any node to select every node in the instance, then drag one of them.
Rotate the selected instance: hold Alt and scroll.
Shift-click nodes to add them to the selection; drag on empty canvas to draw a selection box.
Right-click a node or instance for its context menu — visibility, deletion, track assignment, and more.
Node placement mode (N, or View ▸ Node Placement Mode) is for clicking nodes down one at a time rather than dragging a pre-posed template.

Visibility¶

Nodes that are genuinely occluded should be marked not visible rather than parked somewhere plausible — a guessed point is training signal that says the part is there, which it isn't.

To mark a node non-visible, right-click it ▸ Mark Node Non-Visible (or Mark Node Visible to undo it).
V toggles View ▸ Show Non-Visible Nodes — whether non-visible nodes are drawn at all. If the pointer is over a node, or nodes are selected, V also flips those nodes' visibility.

View ▸ Display controls how much of the hidden geometry you see:

Mode	Shows
Manual	Whatever you toggled by hand
Only selected (with hidden points)	Just the selected instance, hidden points included
All instances, visible points only	Every instance, visible points only
All visible + selected hidden points	Everything, plus the selected instance's hidden points

Selecting¶

Action	Shortcut
Select next instance	`
Select next / previous instance and zoom to it	`Shift`+`Down` / `Shift`+`Up`
Clear selection	`Esc`

Clicking an instance on the canvas selects it; the Instances panel lists the current frame's instances with their track, type, and score.

Copy and paste¶

Action	Shortcut
Copy instance	`Cmd`+`C` / `Ctrl`+`C`
Paste instance	`Cmd`+`V` / `Ctrl`+`V`
Copy instance track	`Cmd`+`Shift`+`C` / `Ctrl`+`Shift`+`C`
Paste instance track	`Cmd`+`Shift`+`V` / `Ctrl`+`Shift`+`V`

Copying a track copies the identity assignment rather than the pose — see Tracks.

Deleting¶

Action	Shortcut
Delete selected instance	`Cmd`+`Backspace` / `Ctrl`+`Backspace`
Delete instance and its track	`Cmd`+`Shift`+`Backspace` / `Ctrl`+`Shift`+`Backspace`
Delete predictions in a dragged area	`Cmd`+`K` / `Ctrl`+`K`

The Labels menu has the bulk options:

Menu item	Deletes
Delete Predictions on Current Frame	Every prediction on this frame
Delete Predictions…	Predictions matching one filter: score threshold, frame range, on user-labeled frames, max instances per frame, track, or instance type
Delete Predictions from Area… (`Cmd`+`K` / `Ctrl`+`K`)	Predictions inside a rectangle you drag
Delete All Predictions…	Every prediction in the project, after a confirm

Working with predictions¶

Predicted instances are drawn distinctly from user labels (View ▸ Color Predicted Instances changes how). They are suggestions until you accept them.

Action	How
Accept one prediction	Double-click it
Accept all on this frame	`Cmd`+`Shift`+`A` / `Ctrl`+`Shift`+`A`, or Labels ▸ Accept All Predictions on Current Frame
Accept all in the project	Labels ▸ Accept All Predictions

Accepting converts a PredictedInstance into a user-labeled Instance — which is what training actually consumes.

Negative frames¶

Labels ▸ Mark Frame as Negative records that a frame genuinely contains no animals. This is real training signal, not an empty frame that you skipped, and the label quality check will flag a negative frame that still has instances on it.

Undo¶

Every edit goes through a command with a frame-level snapshot, so Cmd+Z / Ctrl+Z and Cmd+Shift+Z / Ctrl+Shift+Z work across the whole editing surface — not just node drags.

Hints¶

Labels ▸ Show Hints During Labeling puts contextual tips on screen as you work. Help ▸ Labeling Tips… opens the full list any time.