Skip to content

Object tracking

An Object session finds subjects in footage (people, cars, and anything else the detection model knows) and follows them as named targets. Use it for roto guides, blur and mask targets, crowd counting, or attaching graphics to moving subjects.

Get a model first

Object detection runs on your machine with a YOLO model. YOLO is a family of fast image-detection neural networks; the model is a single ONNX file Kitly runs locally, so no footage leaves your computer.

  • The quick route: click Download recommended model (YOLOX-Nano, Apache-2.0) right in the session.
  • The custom route: register any YOLOv8 or YOLOX ONNX model under Settings, via the custom models list. See Settings. A model trained on your own classes (props, creatures, logos) works here too.

No model, no detections

Detection buttons won't do anything useful until a model is downloaded or registered. If the session seems inert, this is why.

Detect

Two detection actions:

  • Detect Objects runs on the current frame only and shows transient boxes. Good for a quick "does the model see it" check.
  • Detect Video runs the whole clip as a job and caches every frame's detections. Cached detections draw as gray boxes, with a "Show detections" toggle. Cancelling partway still keeps the partial cache.

The class filter chips above the frame show what was found and how many, like "person (342)" or "car (87)". Click chips to narrow which detections are visible and clickable.

Create targets

A detection is just a box; a target is a subject you're following. Ways to promote detections into targets:

  • Click a detection box. It becomes a named target ("Person 1", "Car 2"), renameable in the inspector. If a Detect Video cache exists, the target auto-tracks across the whole clip immediately.
  • Lasso. Drag a box-select over several detections to create targets in bulk.
  • Select All (frame) or Ctrl+A targets everything detected on the current frame.
  • Add filtered as targets turns the current class-filter results into targets, like every "person" in the shot.

How tracking behaves

Targets follow their subject using motion prediction. When the detector loses the subject for a few frames (a lamppost passes in front), the target coasts on its predicted path, drawn as a dashed box, for up to 12 missed frames. If the subject doesn't reappear by then, the target fails out in that direction. The Curves tab plots one confidence curve per target, so dropouts are easy to spot.

Fix broken tracks

Real footage breaks tracks: occlusions split one person into "Person 3" and "Person 7". Two repair tools handle the common cases.

  • Merge. Select two or more same-class targets (Ctrl click their rows, or lasso them) and click Merge. Fragments of the same subject combine into one target; the earliest fragment keeps the name.
  • Fill gaps. Interpolates across internal gaps in a target's track. Interpolated frames draw as dotted boxes at reduced confidence, so you always know which boxes are real detections and which are in-betweens.

A practical cleanup pass: run Detect Video, add your targets, play the clip watching for dashed dropouts, merge the fragments, then fill the remaining small gaps.

Export

Targets export as After Effects Keyframe Data (Position, plus Scale from the box size) or a Blender empties script. Details in Exporting tracks.