PoseTracker API logoPoseTracker

New · Beta

PoseTrackerNet: our own pose estimation model

PoseTrackerNet is PoseTracker's own on-device human pose estimation model for mobile apps. It returns 39 body keypoints in real time on everyday Android phones and iPhones, and is free to use in apps under the PoseTrackerNet licence.

December 1, 2026

Proprietary model. Free to use in apps. Runs on the device.

Why PoseTracker needed its own model

PoseTracker has run on MoveNet since day one, and it still does. MoveNet is a solid model, but it stops at 17 keypoints and at the ankle. Fitness and rehab apps keep asking us for what happens below the ankle and between the shoulders and the hips: the foot, the trunk, the neck.

They also ask for the same thing on every phone their users own, not only on this year's flagship. So we built a model for that: more of the body, at the camera's frame rate, on everyday phones.

What we measured so far

These are early measurements on a pre-release model. Every figure comes with its conditions, and all of them will be replaced by the release figures.

What we measuredResultConditions
Time per frame, model alone, on a Samsung Galaxy A13PoseTrackerNet: 10.1 to 10.8 ms per frame. MoveNet Lightning: 20.0 to 24.3 ms. That is 1.9 to 2.4 times less time.Pre-release model, 8-bit files on both sides, 4 CPU threads, synthetic frames, three rounds, 17 November 2026
Live 30 fps camera for 10 minutes, on the same Galaxy A13PoseTrackerNet held 29.6 frames per second, with a model time of 13.2 to 13.4 ms against 20.5 ms for MoveNet Lightning. MoveNet Lightning also held the camera rate.Pre-release model, 8-bit, 4 CPU threads, phone charging, 17 November 2026
240 fps camera over 60 seconds, on an iPhone 18 ProPoseTrackerNet followed the 240 fps camera at 239.9 frames per second.Pre-release model, 8-bit, 4 CPU threads, 17 November 2026
Accuracy on the 17 shared COCO joints: PCK@0.1 on COCO val 2017 (991 single-person crops)PoseTrackerNet 0.920. MoveNet Lightning 0.928. MoveNet Thunder 0.944.Pre-release 8-bit model against MoveNet float files, any score, 17 November 2026

A first version.PoseTrackerNet is the first version of our own proprietary model. On the 17 shared COCO joints it is slightly less accurate than MoveNet Lightning (PCK@0.1 of 0.920 against 0.928 on COCO val 2017, pre-release 8-bit model against MoveNet float files, 17 November 2026), and it takes 1.9 to 2.4 times less time per frame (model alone, in 8-bit, on a Samsung Galaxy A13, 4 CPU threads, synthetic frames, three rounds, 17 November 2026). We keep training it, and we expect accuracy to improve in later versions.

Where MoveNet is ahead. MoveNet Thunder is more accurate than PoseTrackerNet on the 17 shared joints, and MoveNet Lightning is very slightly ahead as well. PoseTrackerNet's advantage is time per frame in 8-bit and the number of keypoints, not placement accuracy.

What is not measured yet. The 22 additional keypoints have not been measured against human annotation. We will publish the full benchmark, with its protocol, at release.

Preliminary figures, measured on a pre-release model on 17 November 2026. They will be replaced at release.

From 17 keypoints to 39

keypoints

  • Feet
  • Neck
  • Trunk
  • Shoulder line

MoveNet · 17

PoseTrackerNet · 39

Illustration. The exact keypoint map is published with the release.

PoseTrackerNet returns 39 keypoints for one person, in 2D: the 17 COCO keypoints that MoveNet returns, plus points on the feet, the neck, the trunk and the shoulder line. That is what lets an app look at a heel leaving the floor, a rounded back or a head position, with one model.

What is free and what is paid

FreePaid, as today
PoseTrackerNet, the modelThe PoseTracker engine
The PoseTracker SDK and iframeRep counting, joint angles, form score
Body keypoints, on the deviceExercise catalog and the Assistant for custom movements

PoseTrackerNet is proprietary, and free to use in apps, including commercial ones, under the PoseTrackerNet licence. The licence text is published with the release. See pricing

How you will use it

The recommended way is the PoseTracker SDK. It handles image preparation, keypoint decoding, person tracking and smoothing, so you get stable keypoints instead of raw model outputs. If you already use PoseTracker, you will switch models without rewriting your integration.

Read the quickstartHow on-device pose estimation works

Licensing for large teams

Teams that want to own their stack can license the model alone, the model with the PoseTracker engine, or a custom model trained on their own data.

Talk to us

PoseTrackerNet is built by Movelytics

What PoseTrackerNet is not

  • One person at a time.
  • 2D only: no depth, no 3D.
  • Wrists, but no hand keypoints.
  • Not a medical device.
  • Not open source: a proprietary model under licence.

Questions about PoseTrackerNet

Short answers. Each one stands on its own.

What is PoseTrackerNet?

PoseTrackerNet is PoseTracker's own on-device human pose estimation model for mobile apps. It returns 39 body keypoints in real time on everyday Android phones and iPhones, and is free to use in apps under the PoseTrackerNet licence. It is the model behind the keypoints. PoseTracker is the SDK, API and engine built around it.

Is PoseTrackerNet open source?

No. PoseTrackerNet is a proprietary model. It is free to use in apps, including commercial ones, under the PoseTrackerNet licence, which is published with the release.

Is PoseTrackerNet free?

Yes, the model, the PoseTracker SDK and body keypoints are free. The PoseTracker engine on top (rep counting, joint angles, form score) is a paid product, as it is today.

How many keypoints does PoseTrackerNet return compared with MoveNet?

39 against 17. PoseTrackerNet returns the 17 COCO keypoints plus points on the feet, neck, trunk and shoulder line, for one person, in 2D.

How fast is PoseTrackerNet compared with MoveNet Lightning?

Model alone, in 8-bit, PoseTrackerNet takes 10.1 to 10.8 ms per frame on a Samsung Galaxy A13, against 20.0 to 24.3 ms for MoveNet Lightning: 1.9 to 2.4 times less time (pre-release model, 8-bit files on both sides, 4 CPU threads, synthetic frames, three rounds, 17 November 2026). With a live 30 fps camera on that phone, both models hold the camera rate.

Is PoseTrackerNet as accurate as MoveNet?

On the 17 shared joints it is close to MoveNet Lightning (PCK@0.1 of 0.920 against 0.928 on COCO val 2017, pre-release 8-bit model against float files). MoveNet Thunder remains more accurate, at 0.944.

Does the video leave the phone?

No. The model runs on the device and the camera feed is not uploaded.

When will PoseTrackerNet be publicly available?

Public release is planned by December 1, 2026. Beta testers get access first.

How do I get access before the public release?

Join the beta list. We open access through the PoseTracker SDK and email you when yours is ready. Join the beta list →

Will PoseTracker still support MoveNet?

Yes. PoseTracker runs on MoveNet today and MoveNet stays available. PoseTrackerNet is an additional model you can choose.

Can I license the model or get a custom model?

Large teams can license the model alone, the model with the PoseTracker engine, or a custom model trained on their own data. Talk to us →

Be among the first to build with PoseTrackerNet.

Join the beta list