Research note · Benchmark
Real-time pose on a low-end Android phone: what we measured
On 17 November 2026, an internal build of the PoseTracker Model held 28.7 frames per second with the camera on a Huawei FIG-LX1 2017 (Android 9) over 10 minutes, against 21.4 for MoveNet Lightning in 8-bit on the same phone. Alone, the model takes 18.4 ms per frame. These figures are preliminary and measured on an internal build. This note explains how they were obtained and what they do not prove.
Last updated: By the Movelytics team
01Question
The question: does it hold on the phone nobody tests on?
Pose models are usually presented with a speed measured on a recent flagship phone, for a few seconds. That number is true and nearly useless. The people who use a fitness or rehabilitation app do not all own a recent flagship, and nobody exercises for five seconds.
The question a product team needs answered is narrower and harder: on a low-end phone, with the camera running, after several minutes, how many frames per second are actually processed? That is what we set out to measure.
We chose a Huawei FIG-LX1 2017 running Android 9. It is an entry-level Android phone that is several years old: the kind of device that is still in use and that a benchmark rarely includes.
02Protocol
Two tests, fixed before the first run
We fixed the protocol before measuring, so that the results could not shape the method.
Test 1: the model alone
The model runs on synthetic frames, with no camera, for 30 seconds on 4 processor threads. We record the median time per frame. Each model is run in 3 rounds, in a different order each time, with 2 minutes of rest between rounds so that heat from one run does not penalise the next. We report the three medians, and a ratio between two models is always taken inside one round.
"Model alone" isolates what the model itself costs. It is the right figure to compare two models. It is the wrong figure to promise to an app.
Test 2: with the camera, for ten minutes
The camera is held at 30 frames per second and the model processes its frames for 10 minutes. We report the frames per second actually processed over the whole run, and the slowest 10-second window, because a user feels the worst moment, not the average.
We compared the build with the public MoveNet Lightning files, in 8-bit and in floating point. They are the natural reference: small, public, and widely used on phones.
03Results
What the phone gave us
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
Alone, the build takes 18.4 ms per frame. MoveNet Lightning takes 26.2 ms in 8-bit and 38.6 ms in floating point. The three rounds agree closely, which tells us the rest periods did their job.
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
With the camera, over 10 minutes, the build processes 28.7 frames per second, with a slowest window at 22.1. MoveNet Lightning processes 21.4 in 8-bit and 14.1 in floating point.
04Reading
What the two tests say together
Three things stand out.
- The order holds. The model that is faster alone is also faster with the camera. The model-alone test is a fair way to rank models.
- The gap to the camera rate is the real story. The camera delivers 30 frames per second. The build processes 28.7 of them. The floating-point public model processes 14.1: fewer than half. On this phone, that is the difference between a skeleton that follows the body and one that lags behind it.
- Model-alone time is not app time. A model that takes 18.4 ms alone could in theory process many more frames per second than the camera delivers. With the camera it does not keep that margin, because the phone heats up, and because processor cores slow down between two frames when the work arrives at the camera's rhythm. Anyone who promises an app the model-alone figure will be wrong.
The slowest window matters as much as the average. At 22.1 frames per second in its worst ten seconds, the build is still usable for counting repetitions. We would not call that comfortable, and we report it for that reason.
05Limits
What this does not prove
- One phone, one unit. Another Huawei FIG-LX1 2017 could behave differently, and another low-end model certainly will.
- The order of the camera runs favours us. The three ten-minute runs followed each other. The build ran first, on the coolest phone; MoveNet Lightning in 8-bit ran last, on the warmest. The runs have to be repeated in the reverse order.
- The runtime settings were chosen for our files. A configuration tuned for MoveNet could narrow the gap.
- Speed says nothing about accuracy. On a well-framed person the build and MoveNet Lightning are level, and MoveNet Thunder is more accurate than both. The benchmarks page gives the accuracy figures and a section on where the build is behind.
- An internal build is not a product. The model that ships may be faster or slower.
06Next
What we measure next
The same series on more phones, with more than one unit where we can. The camera runs in both orders. Steadiness from frame to frame on video, which no figure in this note captures and which users see immediately. And standard COCO keypoint AP next to PCK, so that our accuracy figures can be set against published ones.
The method will stay the same: the device, the duration, the date and the status next to every number. It is the method described on our technology page.
Private beta
Measure it on the devices your users hold.
The PoseTracker Model is in private beta.