PoseTracker API logoPoseTracker

Introducing PoseTrackerNet, PoseTracker's own pose estimation model

By Fabrice Sepret3 min read

Today we are announcing 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.

Why now

PoseTracker has run on MoveNet since day one, and it still does. The requests we get most often are for parts of the body MoveNet does not return: the foot below the ankle, the trunk between the shoulders and the hips, the neck. PoseTrackerNet returns the 17 COCO keypoints plus points on the feet, neck, trunk and shoulder line: 39 in total, for one person, in 2D.

What we measured so far

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 for 10 minutes on the same Galaxy A13, PoseTrackerNet held 29.6 frames per second with a model time of 13.2 to 13.4 ms, against 20.5 ms for MoveNet Lightning (pre-release model, 8-bit, 4 CPU threads, phone charging, 17 November 2026). MoveNet Lightning also held the camera rate.

On an iPhone 18 Pro, PoseTrackerNet followed a 240 fps camera at 239.9 frames per second over 60 seconds (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).

This is the first version of our own proprietary model. It is slightly less accurate than MoveNet Lightning on the 17 shared joints, as the figures above show, and its model time per frame is 1.9 to 2.4 times shorter on that Galaxy A13 in 8-bit. We keep training it, and we expect accuracy to improve in later versions.

MoveNet Thunder remains more accurate, and the 22 additional keypoints have not yet been measured against human annotation. These figures come from a pre-release model and will be replaced at release, together with the full benchmark and its protocol.

What stays the same

The model runs on the device and the camera feed is not uploaded. The PoseTracker SDK remains the way to integrate, and MoveNet stays available. The PoseTracker engine (rep counting, joint angles, form score) stays a paid product; the model, the SDK and keypoints are free.

What it is

PoseTrackerNet is a proprietary model under licence, not an open-source release. The licence text will be published with the model.

Get early access

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

Need help building your fitness solution?

Want to learn how PoseTracker can enhance your fitness app? Contact us for a demo and see how easy it is to bring real-time tracking and personalized feedback to your clients.

BOOK A CONSULTATION

Or contact us :

— by email : contact@posetracker.com

— by phone or : +33 6 45 20 01 91

Try the Assistant · Fitness SDK guide · SDK overview