Link Search Menu Expand Document

ossia

Point Tracker

Point Tracker 2D and Point Tracker 3D associate detections across frames, smooth their motion and maintain identities through short occlusions. They do not detect objects in an image: feed them point detections from computer vision, OSC or a script.

Input and output

Points accepts a list of vectors, coordinate sublists, or maps containing a position and optional confidence. A flat numeric list is grouped by two coordinates in 2D or three in 3D. A third component of a 2D detection means confidence; a fourth component of a 3D detection means confidence. Use consistent coordinate units: 3D positions are not automatically normalized.

Tracks emits records containing id, slot, state, creation_time, age, time_since_seen, position, position_raw, velocity, confidence, provisional and reacquired. Positions, Ids and Count provide simpler views; Entered, Confirmed, Exited and Revived report lifecycle events. Tracks can connect directly to Entity To MIDI.

Settings workflow

  1. Tune Motion Gate, Max Speed, Position Noise and Motion Noise to the units and plausible motion of the source.
  2. Set High Confidence, Low Confidence and New Track Confidence. Two-Stage Association can use lower-confidence detections to sustain existing tracks without creating new ones.
  3. Balance Confirm Time, Confirm Hits and Confirm Window against onset latency. Emit Unconfirmed exposes provisional tracks immediately; disable it for a more conservative output.
  4. Set Coast Time for temporary disappearance and Revive Time for reacquisition. Smooth, Min Cutoff, Beta, Prediction Lead and Output Deadband balance jitter against response time.
  5. For a fixed voice bank, set Slot Count, Allocation, Steal Policy and Hold Time. Slots are reusable; an id identifies a track, not a permanent hardware channel. Choose compact or slot-oriented Data Format and Order By as needed.

Reset IDs forgets all tracks and restarts numbering. Identity can still become ambiguous when objects cross or detections disappear; tune against real input rather than treating tracking as guaranteed recognition.