Metric ranked #1: three VBT apps tested against GymAware (Li et al., 2024)

Three VBT apps went into an independent lab at South China Normal University. One came out on top.

Researchers ran Metric (computer vision), Tempro (phone-mounted IMU) and Spleeft (Apple Watch IMU) simultaneously against a GymAware linear transducer on the bench press. Metric finished first — and not by a nose. r = 0.937 against a $2,000 tether unit, with the two accelerometer-based apps left in the 0.73–0.75 range.

Read the full paper →

Li, A., Tang, J., & Feng, D. (2024). Reliability and Validity of Mobile Device Apps for Measuring Barbell Velocity During the Bench Press. In 2024 International Conference on Sports Technology and Performance Analysis (ICSTPA 2024), December 13–15, 2024, Guangzhou, China. ACM, pp. 417–422. https://doi.org/10.1145/3723936.3724001

Study overview

Fifteen strength-trained men (age 21.0 ± 2.8 years, body mass 75.0 ± 11.8 kg, height 178.1 ± 5.5 cm, 2–5 years of training) completed a 1-RM bench press test, then five concentric bench press reps at 50% of that 1-RM — 75 reps in total. Every rep was recorded simultaneously by four systems:

• GymAware LT (the reference), sampling at 1000 Hz, running GymAware software 5.0.0
• Metric on an iPhone 15, 720p at 60fps, tripod-mounted within a 0–25° angle from the end of the barbell
• Tempro on an iPhone XS Max, using the phone’s built-in accelerometer, attached to the bar
• Spleeft on an Apple Watch S9, worn on the participant’s forearm

Mean concentric velocity was the only variable compared. Participants were told to lift as fast as they could, with 15 seconds between reps and 4 minutes between sets.

Metric finished first

Pearson’s r = 0.937 against the GymAware, which the authors describe as a “practically perfect” correlation, with an ICC of 0.935 (95% CI 0.898 to 0.958) and a standard error of estimate of 0.017 m/s — the tightest of the three apps.

The two IMU-based apps weren’t close:

App	Pearson r	ICC (95% CI)	Bias vs GymAware
Metric (computer vision)	0.937	0.935 (0.898–0.958)	+0.064 m/s
Tempro (phone IMU)	0.751	0.702 (0.567–0.801)	−0.055 m/s
Spleeft (watch IMU)	0.733	0.713 (0.581–0.809)	−0.070 m/s

Here’s the rep-by-rep data, redrawn from the paper’s own figure. A phone on a tripod against a $2,000 tether unit, rep for rep:

An SEE of 0.017 m/s is the number to sit with. That is the typical gap between a camera pointed at a barbell and a 1000 Hz linear transducer physically bolted to it — 17 thousandths of a metre per second, on hardware most lifters already own.

All three apps showed a statistically significant systematic bias (p < 0.001 on paired-samples t-tests). Metric read slightly high relative to the GymAware (+0.064 m/s, 95% CI 0.060 to 0.068); Tempro and Spleeft both read low.

The bias is the interesting part

An ICC of 0.935 with a fixed +0.064 m/s offset is a specific kind of result: the rep-to-rep tracking is tight, but the whole distribution sits a fraction above the tether unit’s.

The authors’ explanation is a start-of-rep detection difference — the tether registers barbell movement fractionally before the vision system does, so GymAware’s timing window includes more of the slow initial motion off the chest, dragging its mean velocity down. Metric starts the clock slightly later, on faster motion, and returns a marginally higher mean. It’s a definitional gap in where a rep begins, not a tracking failure, and it’s consistent with what other computer-vision validation work has reported.

Practically: a fixed offset of this shape doesn’t affect velocity loss percentages, set-to-set comparisons, or fatigue monitoring, because it cancels within a session. It does matter if you’re pulling velocity thresholds straight out of a paper built on tether data and applying them to app numbers without recalibrating to your own device.

Worth noting the magnitude in context. The reps in this dataset ran roughly 0.43–0.69 m/s on the GymAware, averaging 0.56 m/s, so a 0.064 m/s offset is around 11% of a typical reading — small in absolute terms, not nothing in relative terms. Our own June 2026 internal validation against a GymAware unit — 95 reps across bench, deadlift, squat and cleans on iOS v6.9 and Android v1.2 — returned a mean velocity bias of +0.02 m/s with CCC 0.982. The paper does not report which version of Metric was tested, but the data was collected before December 2024, so it predates the v6.9 accuracy work.

Vision vs inertial sensors

The clearest finding in the paper isn’t about Metric specifically — it’s about the measurement approach. Both IMU apps landed in the moderate range (ICC 0.70–0.71), and the authors are blunt about it: Tempro and Spleeft “are not effective and reliable measures of barbell velocity.”

Their reasoning is the low sampling frequency of consumer-device accelerometers, plus the need to integrate acceleration over time to get velocity — an operation that accumulates error. A vision system measures displacement directly and differentiates it, which is the more forgiving direction to go.

One methodological caveat the authors flag: the Spleeft watch was worn on the forearm rather than mounted on the bar, so its readings pick up lateral wrist movement that the bar-mounted and vision systems don’t. That’s a fair-test problem, though it’s also how the product is actually used.

Limitations

The paper is honest about its scope, and three limits matter when reading it:

1. One load only. Everything was measured at 50% 1-RM. Accuracy at heavier, slower loads — where absolute errors matter more as a percentage of the reading — wasn’t tested. The Grossi paper covers the opposite end of that range at 1-RM.
2. Mean velocity only. No peak velocity and no mean propulsive velocity, both of which are widely used in VBT programming.
3. One exercise, 75 reps. Bench press only, fifteen participants, single session.

What to take from this

Independent researchers, hardware we had no hand in, an analysis we had no input on — and Metric came out ahead of every other app on the bench. That’s the result we’re proudest of here. Not “an app did okay against lab gear”, but the camera in your pocket tracked the bar more faithfully than dedicated inertial sensors did, with an SEE of 0.017 m/s against a 1000 Hz transducer.

If you already use Metric, the useful detail is the direction and size of the offset: Metric reads a touch fast against a tether, because of where each system decides a rep has begun. It’s consistent across reps, it’s explainable, and it doesn’t disturb the within-session comparisons that most VBT work is built on.

If you’re still deciding what to track with, this is the cleanest head-to-head available: same lifters, same bar, same reps, four systems recording at once, one clear winner.

The full paper is in the ACM Digital Library at the link above.