Product · Our own modelPrivate beta
PoseTracker Model: real-time pose, built for the phones people own
The PoseTracker Model is Movelytics' own human pose estimation model. It is designed for mobile processors, returns 39 body keypoints where public models of the COCO family return 17, and runs with 8-bit inference. It ships first inside PoseTracker. It is in private beta: its benchmarks are preliminary and measured on an internal build.
Last updated:
Forward fold · the 39 exported keypointsIllustration. Authored pose sequence, not live inference.
01Problem
The public models stopped moving. Phones did not get faster for everyone.
PoseTracker was built on public pose models. They are good, and they made real-time tracking possible on a phone. But they have not changed in years, and they were not tuned for the low-end Android phones that many people use today.
Our customers ask for two things those models do not give together:
- Real time on the device, on any phone. Not only on last year's flagship.
- More keypoints, and custom ones. Seventeen points describe limbs. They say little about the feet, the trunk or the spine, which is where posture and mobility are read.
02Why we built it
We are building the reference pose model for mobile.
That is an ambition, stated as one. The way there is to own the model: train it ourselves, choose its keypoints, design it for the processors it will run on, and measure it on the phones our customers' users hold.
There is a business reason too. PoseTracker has historically run on open-source models such as MoveNet, with our own movement-recognition engines on top. Clients who want to own the whole technology can still have exactly that, with the full intellectual property delivered to them. Clients who want a more optimised model with more keypoints can choose ours, under an annual licence. And for everyone who uses the API, owning the model is what keeps it stable and relevant. How the two options compare.
Owning the model also changes what the lab can do for a client. A keypoint that matters for one sport or one rehabilitation protocol can be added. That is not possible with a model someone else froze years ago.
03How it works
39 keypoints, 192-pixel input, 8-bit inference
39
keypoints
Face, body, feet, trunk, neck and shoulders. Private beta; benchmarks are preliminary and measured on an internal build.
18.4ms
per frame, model alone
Huawei FIG-LX1 2017 (Android 9), 4 threads, 2026-11-17. Private beta; benchmarks are preliminary and measured on an internal build.
28.7fps
with the camera, 10 minutes
Huawei FIG-LX1 2017, camera at 30 fps, 2026-11-17. Slowest 10 s window: 22.1. Private beta; benchmarks are preliminary and measured on an internal build.
- 39 body keypoints
- The 17 points of COCO-style models, plus 22 more: feet, the neck, the shoulders, and trunk points down the front and down the back, which matter for posture and mobility.
- 192-pixel input
- The model reads a small crop around the person. A small input is what keeps the time per frame low on a mobile processor.
- 8-bit inference
- Weights and computations use 8-bit integers. This is what entry-level processors run fastest. See our note on 8-bit inference.
- Two models, one SDK
- A light model for low-end Android and as a universal fallback; a heavier, more precise model for recent phones. The SDK picks per device.
| Group | Count | Keypoints (names in the exported files) |
|---|---|---|
| Face | 5 | nose, left_eye, right_eye, left_ear, right_ear |
| Body | 13 | lower_neck_front; shoulders, elbows, wrists, hips, knees, ankles (left and right) |
| Feet | 4 | left_toe_inner, right_toe_inner, left_heel, right_heel |
| Trunk | 10 | trunk_f1 to trunk_f5 down the front, trunk_b1 to trunk_b5 down the back |
| Neck and shoulders | 7 | lower_neck_back, upper_neck_front, upper_neck_back, left_deltoid, right_deltoid, left_trapezius, right_trapezius |
The figure at the top of this page shows what the extra points add. With 17 keypoints, a forward fold is a hip angle. With 39, the curve of the spine and the position of the feet become measurable. The figure uses the names of the exported model files; the positions are drawn for the illustration, not produced by the model.
04What you get
In the beta
- Real-time tracking on the device, on iOS and Android, including low-end Android phones.
- 39 keypoints instead of 17, in the same call.
- The model inside PoseTracker: the outputs you already use (angles, reps, scores) computed from a richer skeleton.
- A direct line to the team that trains the model, to discuss keypoints your product needs.
The beta is private. We open it team by team, starting with products that need more keypoints or better speed on older phones.
05Proof
What we measured so far
On 17 November 2026 we ran a first benchmark on two phones, next to the public MoveNet files. On the phones tested, the internal build is about 1.4× faster than MoveNet Lightning in 8-bit, with 39 keypoints instead of 17, at comparable accuracy on well-framed people.
| Measure | PoseTracker Model (8-bit) | MoveNet Lightning (8-bit) |
|---|---|---|
| Time per frame, model alone, Huawei FIG-LX1 2017 | 18.4 ms | 26.2 ms |
| Frames per second with the camera, 10 min, Huawei FIG-LX1 2017 | 28.7 | 21.4 |
| Time per frame, model alone, iPhone 18 Pro | 1.8 ms | 2.5 ms |
| Keypoints | 39 | 17 |
We also publish where the build is behind: fine placement at tight tolerance, occluded joints, and the accuracy of MoveNet Thunder on a well-framed person. It is all on the benchmarks page, with the raw numbers as a CSV file.
06Data and privacy
Same rule as every Movelytics product
The model runs on the device. The video is not uploaded for inference. Movelytics collects usage metering only, and the app that integrates the model keeps its users' data. Details in our data commitments and on the technology page.
07FAQ
Questions about the PoseTracker Model
What is the PoseTracker Model?
The PoseTracker Model is Movelytics' own real-time human pose estimation model, designed for mobile processors. It returns 39 body keypoints, uses a 192-pixel input and runs with 8-bit inference. It is in private beta.
How many keypoints does it return?
39: 5 face points, 13 body points, 4 foot points, 10 trunk points (five down the front, five down the back) and 7 neck and shoulder points. That is the 17 points of COCO-style models plus 22 more. The model tracks one person per image, the closest one.
How fast is it?
In a preliminary benchmark of 17 November 2026, an internal build took 18.4 ms per frame alone on a Huawei FIG-LX1 2017 (Android 9, 4 threads) and held 28.7 frames per second with the camera over 10 minutes (slowest 10-second window: 22.1). These figures are preliminary and measured on an internal build, not on the shipped model.
How does it compare with MoveNet?
On the phones tested it is about 1.4× faster than MoveNet Lightning in 8-bit, with 39 keypoints instead of 17, at comparable accuracy on well-framed people. MoveNet Thunder is more accurate on a well-framed person. The benchmarks page gives every figure and its conditions.
How do I get access?
Through the private beta. Describe your app, your platforms and the keypoints you need on the beta sign-up page.
Private beta
Need more keypoints, on slower phones?
Tell us about your app. We open the beta team by team.