Preliminary

These figures come from an internal build, not from the shipped model. Measured on 2026-11-17. They will change.

PoseTracker Model · BenchmarksPreliminary

PoseTracker Model benchmark: measured speed and accuracy on named phones

On 17 November 2026 we measured an internal build of the PoseTracker Model next to the public MoveNet files, on a low-end Android phone and on a recent iPhone. 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. This page gives every figure with its conditions, where the build is behind, and the limits of the method.

Last updated: Status: preliminary

01Summary

Three figures, and what they are worth

1.4×

faster than MoveNet Lightning in 8-bit

Time per frame, model alone, on both phones tested. Private beta; benchmarks are preliminary and measured on an internal build.

28.7fps

with the camera, over 10 minutes

Huawei FIG-LX1 2017 (Android 9); slowest 10 s window 22.1; MoveNet Lightning 8-bit: 21.4 fps. Run order not yet reversed. Private beta; benchmarks are preliminary and measured on an internal build.

39

keypoints, against 17

Accuracy compared on the 17 joints both models share; the other 22 keypoints are not scored here. Private beta; benchmarks are preliminary and measured on an internal build.

The strongest claim we make: about 1.4× faster than MoveNet Lightning in 8-bit on the phones tested, with 39 keypoints instead of 17, at comparable accuracy on well-framed people.

We do not claim more. The build is behind in several places, listed in section 05, and the method has limits, listed in section 06.

02Protocol

How the measurements were made

Devices
A Huawei FIG-LX1 2017 running Android 9 (low-end Android) and an iPhone 18 Pro. One unit of each.
Models compared
An internal build of the PoseTracker Model in 8-bit (39 keypoints, 192-pixel input) and the public MoveNet files: Lightning in 8-bit and in floating point, Thunder in floating point (17 keypoints).
Model alone
The model runs on synthetic frames, with no camera, for 30 seconds per run on 4 processor threads. We report the median time per frame. "Model alone" isolates the model's own cost; it is not the speed an app gets.
Three rounds
Each model is run in 3 rounds, with the models in a different order each time and 2 minutes of rest between rounds. We give the three medians, never one alone. A ratio between two models is taken inside one round.
With the camera
The camera is held at 30 frames per second and each model runs for 10 minutes. We report the frames per second actually processed and the slowest 10-second window.
High-rate camera
On the iPhone 18 Pro, the back camera is asked for 240 frames per second for 60 seconds, to see whether the model follows.
Accuracy
COCO val, 991 persons, the 17 joints annotated by humans that both model families share. Score: PCK, the share of joints placed within 5 % (PCK@0.05) or 10 % (PCK@0.1) of the crop side. Every keypoint counts, whatever its confidence score.
Framing
"Tight": the person fills the crop. "Small": the person is small in the frame. Both are 192-pixel crops; MoveNet Thunder receives the crop enlarged to its own input size.

03Speed

Speed on a low-end Android phone and on an iPhone

Lower time per frame is better; higher frames per second is better. All speed figures: 4 threads, measured on 2026-11-17, internal build.

Time per frame, model alone

Huawei FIG-LX1 2017 · Android 9 · 4 threads · median over 30 s · preliminary

  • PoseTracker Model 8-bit18.4ms

    three rounds: 18.4 · 18.4 · 18.4

  • MoveNet Lightning 8-bit26.2ms

    three rounds: 26.2 · 26.2 · 26.2

  • MoveNet Lightning float38.6ms

    three rounds: 38.6 · 38.7 · 38.0

Lower is better

Frames per second with the camera

Huawei FIG-LX1 2017 · camera at 30 fps · 10 minutes · preliminary

  • PoseTracker Model 8-bit28.7fps

    slowest 10 s window: 22.1 fps

  • MoveNet Lightning 8-bit21.4fps

    slowest 10 s window: 18.1 fps

  • MoveNet Lightning float14.1fps

    slowest 10 s window: 12.5 fps

Higher is better

Time per frame, model alone

iPhone 18 Pro · 4 threads · median over 30 s · preliminary

  • PoseTracker Model 8-bit1.8ms

    three rounds: 1.8 · 1.8 · 2.0

  • MoveNet Lightning 8-bit2.5ms

    three rounds: 2.5 · 2.7 · 3.0

Lower is better

Following a 240 fps camera

iPhone 18 Pro · back camera asked for 240 fps · 60 s · preliminary

  • PoseTracker Model 8-bit239.9fps

Bar scaled to the camera's 240 fps

The same figures as tables

ModelKeypointsms per frameThree rounds (ms)
PoseTracker Model 8-bit3918.418.4 · 18.4 · 18.4
MoveNet Lightning 8-bit1726.226.2 · 26.2 · 26.2
MoveNet Lightning float1738.638.6 · 38.7 · 38.0
Huawei FIG-LX1 2017, Android 9, model alone, 4 threads, median time per frame over 30 s. Private beta; benchmarks are preliminary and measured on an internal build.
ModelKeypointsFrames per secondSlowest 10 s window
PoseTracker Model 8-bit3928.722.1
MoveNet Lightning 8-bit1721.418.1
MoveNet Lightning float1714.112.5
Huawei FIG-LX1 2017, camera held at 30 fps, 10 minutes per model, 4 threads. Private beta; benchmarks are preliminary and measured on an internal build.
ModelKeypointsms per frameThree rounds (ms)
PoseTracker Model 8-bit391.81.8 · 1.8 · 2.0
MoveNet Lightning 8-bit172.52.5 · 2.7 · 3.0
iPhone 18 Pro, model alone, 4 threads, median time per frame over 30 s. Private beta; benchmarks are preliminary and measured on an internal build.

04Accuracy

Accuracy on the joints both models share

The PoseTracker Model returns 39 keypoints and MoveNet 17. Accuracy can only be compared on what both have: the 17 body joints, on COCO val (991 persons). Figures are PCK@0.05 / PCK@0.1; higher is better.

ModelKeypointsPCK@0.05PCK@0.1
PoseTracker Model 8-bit390.8320.920
MoveNet Lightning float170.8380.928
MoveNet Thunder float170.8750.944
COCO val, 991 persons, 17 shared joints, tight framing (the person fills the crop). Private beta; benchmarks are preliminary and measured on an internal build.
ModelKeypointsPCK@0.05PCK@0.1
PoseTracker Model 8-bit390.8330.916
MoveNet Lightning float170.7970.889
COCO val, 991 persons, 17 shared joints, small person in the frame. Private beta; benchmarks are preliminary and measured on an internal build.

Reading. On a well-framed person, the build and MoveNet Lightning are level: 0.832 / 0.920 against 0.838 / 0.928, Lightning slightly ahead. MoveNet Thunder is ahead of both at 0.875 / 0.944. When the person is small in the frame, the build is ahead of Lightning: 0.833 / 0.916 against 0.797 / 0.889.

The small-person comparison favours us: MoveNet is meant to run with its own crop tracking from frame to frame, which brings it closer to tight framing in real use. Its small-person figures understate it.

05Where we are behind

Where the build is behind

We looked for the places where the build loses, on purpose. Three stand out.

Fine placement at tight tolerance
When the tolerance is tightened, the build finds a joint about as often as MoveNet but places it less finely. To the eye, the skeleton looks slightly loose. MoveNet Lightning and MoveNet Thunder both place joints more precisely at the tightest tolerance.
Occluded joints
For joints that are annotated but hidden in the image, the build is behind both MoveNet Lightning and MoveNet Thunder.
MoveNet Thunder on a well-framed person
Thunder reaches 0.875 / 0.944 where the build reaches 0.832 / 0.920. Thunder is a larger model with a larger input; on the low-end phone it is several times slower. The build does not match its placement.

On a well-framed person, MoveNet Lightning in floating point is also slightly ahead of the build (0.838 / 0.928 against 0.832 / 0.920). "Comparable accuracy" in our headline means exactly that: level, not better.

06Limitations

Limits of this benchmark

  • One unit per phone. One Huawei FIG-LX1 2017 and one iPhone 18 Pro. Another unit of the same model could give different figures.
  • Thermal order of the camera runs. On the Huawei FIG-LX1 2017, the three camera runs followed each other. The build ran first, on the coolest phone, and MoveNet Lightning in 8-bit ran last, on the warmest. This favours the build. The runs have to be repeated with the order reversed.
  • Model alone is not app speed. Model-alone figures use synthetic frames. With the camera running, the time per frame is higher: heat, and processor cores that slow down between two frames.
  • Headline claims use only the set both models share. The accuracy comparison rests on COCO val and on the 17 shared joints. The extra keypoints of the build (feet, trunk, spine) have no human annotation in that set and are not scored here.
  • Runtime settings. The app's runtime settings were chosen for our 8-bit files. Another runtime configuration could suit MoveNet better.
  • Accuracy was computed off the phone. The same model files were scored on a desktop computer. A model file gives the same keypoints on a phone up to rounding; this was not verified on the phone itself.
  • Not measured yet. Steadiness from frame to frame on video, behaviour on the phone's own camera images, and standard COCO keypoint AP next to PCK.
  • Preliminary build. The figures describe an internal build, not the model that will ship.

07Data

Download the figures

Every number on this page is rendered from one data file. The CSV below contains the same values, with the device, the condition, the status and the date on each row.

Download the CSV

Status: preliminary · internal build, not the shipped model · measured on

Private beta

Test it on your own phones.

The beta is the way to check these figures on the devices your users hold.