Performance

Pose detection is one of the heavier things you can run on a phone. These settings have the biggest effect, roughly in order.

1. Use the lite model

pose_landmarker_lite.task is several times faster than heavy and is accurate enough for most live use. Save the bigger models for still images.

2. Limit the frame rate with fpsMode

Most use cases (rep counting, posture checks, skeleton overlays) look smooth at 10–15 detections per second. Processing every 30 or 60 fps frame mostly burns battery.

usePoseDetection(callbacks, RunningMode.LIVE_STREAM, 'pose_landmarker_lite.task', {
  fpsMode: 15, // or 'none' to process every frame (default)
});

This uses VisionCamera’s runAtTargetFps under the hood, so skipped frames never reach MediaPipe.

Built-in throttling on iOS

On iOS, the native module also caps result events to about 15 per second to keep memory stable during long sessions. Android delivers every result, so setting fpsMode matters more there.

3. Detect only as many people as you need

numPoses defaults to 1. Raising it makes each frame slower even when only one person is in view.

4. Leave segmentation masks off

shouldOutputSegmentationMasks defaults to false. Keep it that way unless you need the masks.

5. GPU vs CPU

Delegate.GPU is the default and the fastest on most devices. If you see crashes or memory warnings on older or low-end phones, try Delegate.CPU.

6. Keep work out of onResults

onResults runs on the JS thread. Heavy work there, or a setState that re-renders a large tree on every frame, will make the UI stutter. Do the minimum, and use shared values or refs for anything that updates every frame.

Recommended starting point

{
  numPoses: 1,
  minPoseDetectionConfidence: 0.5,
  minPosePresenceConfidence: 0.5,
  minTrackingConfidence: 0.5,
  delegate: Delegate.GPU,
  fpsMode: 15,
}