How accurate is Atlas? The honest answer — no email required.
Most camera-based movement tools ask for your email before showing you an accuracy number, if they show one at all. We're not going to do that. This page is the real methodology behind every red, yellow, and green you see from Atlas: what's measured well, what's deliberately not reported, where every band edge came from, and the honest limitations of a single 2D camera. Nothing here is gated.
Compared to what?
The realistic alternative for a telehealth visit isn't a $150,000 motion-capture lab. It's a clinician watching you move over a video call and writing "improved" or "not improved" in a chart, from memory, with no way to compare today's movement to last month's frame-by-frame.
Against that baseline, objective, repeatable, timestamped measurement — captured the same way on every visit — is a real improvement, and it doesn't require overstating what a single camera can do. So we won't.
The claims we can back up
Lands in roughly the same error band as clinical goniometry itself — about 2–5°. Goniometry's own published inter-rater error runs 1.9–4.6° (Gallego-Izquierdo 2020 and related literature), and its inter-rater agreement falls as low as 0.16–0.35 ICC in some joints. Atlas is not claiming to beat a skilled clinician with a goniometer — it's claiming to be in the same range, with zero drift between raters, because there is no rater.
The strongest measurement claim on this page. Markerless walking-speed measurement reaches ICC 0.97–0.98 against instrumented reference systems in the published literature. Gait speed also happens to be one of the best-established clinical indicators there is — often called the "sixth vital sign" — with a known minimal clinically important difference of about 0.1 m/s. We measure it with enough precision to detect that change.
The same protocol, the same math, every time — no rater fatigue, no rater-to-rater drift. That makes Atlas best suited to tracking change in one patient over time under a fixed setup, which is exactly how it's used: a baseline, then re-checks against that same baseline.
Photographic posture measurements — the kind Atlas uses for forward-head posture, shoulder-height asymmetry, and pelvic obliquity — have real, published reliability behind the general method (inter-rater ICC in the 0.84–0.90 range in comparable studies). These are reported as screening indices worth a closer look, never as a diagnosis of what's causing them.
A single camera has real limits. We built around them instead of hiding them.
A camera sees your silhouette in one plane at a time. It cannot see rotation happening toward or away from the lens with any real confidence. Rather than display a number we don't trust, Atlas suppresses these findings — a product decision, not a bug:
- ✕Rotational measurements — hip internal/external rotation, shoulder internal rotation, and similar out-of-plane motions are not displayed as measured degrees. The published literature on single-camera rotational accuracy is not strong enough to support it.
- ✕Ankle joint angles — the ankle is a known weak point for markerless motion capture generally, not just for Atlas. We don't report a specific ankle angle as a graded clinical finding.
- ✕A diagnosis — Atlas identifies movement patterns worth a closer look. It never tells you what condition you have. That's Brian's job, and it requires an actual clinical exam.
- ✕Precision past what the method supports — you won't see a reading like "17.3° of forward head posture" from us. Displaying that kind of false precision on a method with several degrees of real-world error is misleading, so we round and band instead of pretending to a decimal place we can't defend.
We label our own confidence. Nobody asked us to.
Every posture finding Atlas reports carries one of three honesty labels, tracked in our engineering documentation and unchanged between what a clinician sees and what powers the patient-facing screen. We'd rather tell you exactly how confident we are in a number than let you assume it's more validated than it is.
Band edges (green ≥75°, yellow 60–75°, red <60° on our own camera-angle proxy scale) were set from three deliberately-graded reference poses captured on one subject — a real derivation, not a guess, but statistically thin. We publish the exact readings: a neutral pose measured 83.8°, a deliberately mild slouch measured 65.0°, and a deliberate full slump measured 55.7°, with the bands drawn between them. We also publish the failure mode we found: the two sides of the body disagreed by as much as 24° on the most exaggerated pose, which we attribute to trunk rotation during capture — an honest limitation, not smoothed over.
Anchored to a published healthy-population radiographic study (Moharrami 2023, n=134) rather than our own captures, read through a surface camera proxy instead of an X-ray. The gap between "what the study measured" and "what our camera can see" is real and we say so plainly rather than borrow the study's precision for our proxy.
Band edges (2°/5° tolerance) come from Brian Jones, DPT's own clinical screening judgment as a starting point — not yet backed by our own calibration captures, and we say that outright rather than imply a study exists where one doesn't yet. As real capture data accumulates, these move to a calibrated status.
Two findings that appear in some posture screens — kyphosis (a rounded upper back) and rounded shoulders — are not graded by Atlas at all today. The camera doesn't have a reliable way to measure them yet, so instead of guessing, the screen tells you honestly that it isn't a reliable read.
A screening tool a clinician reviews. Not a diagnostic device.
Atlas is provider-facing decision support: every finding is traceable back to the specific video and the specific measurement behind it, so Brian Jones, DPT can independently review the basis for anything Atlas flags before it becomes part of your care. Atlas never makes an autonomous diagnostic call, and it never replaces an in-person or telehealth evaluation by a licensed clinician.
We also won't tell you Atlas is "FDA registered" or "FDA listed" the way some competitors do — registering a business with the FDA is a fee and a form, not an approval, and using that language as a quality signal is misleading regardless of who does it. If that ever changes, we'll say exactly what changed and cite it.
Try the free movement screen — same honesty, real time.
Every band you see in the demo is built on exactly what's described on this page. No email required to see a sample result first, either.