Point Tracker
Track points across video frames with the CoTracker3 neural point tracker.
What it does
Point Tracker follows a set of points across video frames using CoTracker3, a neural point tracker. Given an image (video, camera, or any other moving raster source) and a set of points to watch, it reports where each point has moved to on every frame, along with velocity, a confidence score and whether the point is still visible.
With Query Points unconnected, it tracks points placed interactively in the viewport. Each tracked point keeps a stable ID and an age, the number of frames it has survived, so tracks can build trails, gate effects, or be dropped once they drift out of view.
When to use it
A marker, feature or object point followed through a moving shot, in place of hand-animated position
Per-point motion data (position and velocity) driving another effect: a shape attached to a track, a warp around it, a trigger on appearance or loss
Stable point identities over time (track IDs) rather than a one-off detection per frame
| Parameter | Type | Default |
|---|---|---|
image | OneOf([FieldOf(Raster)]) | — |
query_points | OneOf([Vec2Array, Collection]) | — |
enabled | Boolean | true |
window_length | Number | 16 |
visibility_threshold | Number | 0.500 |
Gotchas
Wire a Vec2Array or a point Collection into Query Points to drive the query points from the graph, for example from a detector node. Leave it unconnected to use viewport-placed points.
Visibility Threshold filters which points count as tracked. A point whose visibility score drops below the threshold is treated as lost even though it still has a last-known position. Lower it if points near the frame edge or under light occlusion drop out too eagerly.
Window Length trades smoothness against speed: a longer sliding window of frames gives steadier tracks but costs more to compute per frame. Start low and raise it only if tracks are jittery.
Worked example
Add a Point Tracker and wire a video or camera image into Image.
Click in the viewport to place one or more query points on the frame.
Play the video. Tracks, Positions, and Count update every frame as the points move.
Wire Positions into another node, such as a shape's transform, to have it follow a tracked point, or read
visibilityfrom Tracks to hide it when the point is lost.