How PosturaScreen Measures Posture

Two photos → 17 anatomical keypoints → geometric ratios and angles → a screening report. The full technical methodology, the validity literature it's built on, and the limitations we tag honestly on every report.

What the AI detects: 17 keypoints

The detection layer uses YOLO26-pose (Ultralytics), an open-source pose-estimation model that returns the 17 keypoints of the COCO keypoint specification: nose, both eyes, both ears, both shoulders, both elbows, both wrists, both hips, both knees, both ankles. The COCO 17-keypoint format is the de facto standard for human-pose estimation in computer vision and is what most published pose-based posture research uses.

For each photo, the model returns the (x, y) pixel coordinates of every detected keypoint together with a per-keypoint confidence score. We use the confidence scores as one input to the photo-quality gate (see below).

How each metric is computed

Every metric is a deterministic geometric calculation. There is no per-user model adjustment and no learned-output black box. The same input keypoints always produce the same metric values.

Front-view metrics (11)

• Head tilt — the angle of the eye-line relative to horizontal.
• Shoulder level — the vertical pixel offset between left and right shoulder keypoints, normalized to body height (in subject-space cm).
• Clavicle angle — angle of the shoulder line relative to horizontal.
• Chest symmetry — horizontal offset between shoulder-midpoint and pelvis-midpoint.
• Arm position — horizontal offset of each wrist relative to its ipsilateral hip.
• Waist angle — angle of the shoulder-line relative to the hip-line.
• Pelvic level — vertical offset between left and right hip keypoints.
• Knee alignment — horizontal offset between left and right knee centers.
• Foot position — angle of the foot-line relative to horizontal.
• Q angle (left, right) — angle between the hip→knee vector and knee→ankle vector for each leg. approx

Side-view metrics (6)

• Forward head angle — angle from shoulder to ear relative to vertical. approx
• Ear–shoulder offset — horizontal pixel offset between ear and shoulder, normalized to subject-space cm.
• Thoracic kyphosis — surface curvature estimate from shoulder-to-mid-back angle. approx
• Lumbar lordosis — surface curvature estimate from mid-back-to-hip angle. approx
• Pelvic tilt (sagittal) — angle of the hip-to-knee vector relative to vertical, as a proxy for anterior/posterior pelvic rotation. approx
• Knee angle — angle at the knee formed by hip-to-knee and knee-to-ankle vectors.

The full per-metric formula reference, the normal-range thresholds, and the gender-specific Q-angle cutoffs are documented on our 17 metrics page.

What "approx" means

Five of the 17 metrics carry an explicit approx tag on the report. These are metrics whose ground truth would normally be measured by radiograph, 3-D motion capture, or sustained-pressure goniometry — none of which a single 2-D photo can replicate. We tag them approx for three reasons:

1. So practitioners do not communicate a 2-D screening estimate to clients as if it were a radiographic measurement.
2. So self-screening users understand the limits of what a phone photo can tell them.
3. So the values remain useful for tracking change over time — the inter-session reliability is what we trade off for the convenience of photo-based measurement, and tracking change is robust to the absolute-value uncertainty.

What we're confident about — and what we're not

Photo-based posture screening has known strengths and known weaknesses. We're transparent about both, because anyone consuming a screening number deserves to know how much weight to put on it.

Where we're confident:

• Head and shoulder positions. Side-view head position relative to the shoulder line, and front-view shoulder-height symmetry, are well-defined geometric quantities. When capture conditions are controlled (consistent camera height, distance, and posture cue), inter-session reliability for the same individual is strong. For tracking change over time — which is the primary use case for screening — these are the metrics with the cleanest signal-to-noise.
• Pelvic and chest symmetry. Left-vs-right offsets are relative measurements and are inherently less sensitive to small differences in camera setup than absolute angles are. A 1 cm left-shoulder-higher reading is robust across the kinds of capture variation you see in practice.

Where we're less confident, hence the approx tag:

• Spinal curvature (thoracic kyphosis, lumbar lordosis). We infer these from surface contour — the angles between shoulder, mid-back, and pelvis keypoints. But surface contour is influenced by clothing, body composition, and breathing phase. None of those appear on a radiograph, which is the gold standard for spinal curve measurement. We report these because they are useful for trending within an individual; we tag them approx because they should not be confused with radiographic values.
• Q-angle, forward-head angle, and pelvic tilt. These are derived from 2-D projections of 3-D structures. When the person's body rotation relative to the camera is not perfectly aligned, the absolute number drifts a few degrees. The right use is trending in the same individual under consistent capture — using change over time as the signal. Comparing absolute numbers between two different people, less so.

If you're a researcher who would like to share a peer-reviewed paper that either supports or challenges our threshold choices, please reach out at support@posturascreen.com. We update the methodology when good evidence comes in. For background reading, PubMed search terms that map to this space include "photographic posture analysis", "craniovertebral angle photogrammetry", "two-dimensional posture screening reliability".

How photo screening compares to other measurement methods

Photo-based screening is one point on a spectrum of posture-measurement methods, each with a different trade-off between accessibility and precision. We're explicit about where a 2-D photo sits, so the numbers are read with the right expectations.

A photo does not compete with a radiograph on precision, and we don't claim it does. What it offers is the opposite trade-off: no radiation, no appointment, no specialist equipment, and a result repeatable enough to track change over weeks or months. The published reliability literature (see references) supports exactly that use — strong repeatability for the same person under consistent capture, which is what a screening-and-tracking tool needs.

How to capture photos for reliable measurement

Because the metrics are geometric, capture conditions matter more than camera quality. The same body can read differently if the camera angle or stance changes between sessions. For results that are comparable over time, control these variables:

• Two views. One front-facing and one side-on (left or right, kept consistent between sessions).
• Camera at hip height, and level. A camera tilted up or down foreshortens the body and can invent or exaggerate an angle. The photo-quality gate rejects tilt beyond ±2°.
• 2-3 meters away, so the whole body fits without a wide-angle lens distorting the edges.
• Plain wall, even lighting. Clutter and backlighting degrade keypoint detection.
• Fitted clothing. Baggy layers hide the shoulder, waist, and hip contours the metrics depend on.
• Relaxed, habitual stance. Standing "to attention" for the photo measures a posture the person does not actually hold. The screening value is in the default stance.

For tracking, consistency beats perfection: the same setup each time makes the change between sessions trustworthy, which is the signal a screening tool is built around.

Photo-quality gate

Before measurement, each photo passes through a quality gate. A photo is flagged for re-take when any of the following holds:

• The mean per-keypoint confidence score is below threshold (low overall detection quality — typically poor lighting or motion blur).
• One or more required keypoints are missing (cropped body, occluded by clothing, hands in pockets).
• The camera tilt — inferred from the angle between the horizon of detected ankle keypoints — is beyond ±2°.

If the photo passes, the report shows a "Photo quality: good" notice. If it fails, the report carries a "Photo quality may affect accuracy" notice so a low-quality capture is never silently trusted.

What PosturaScreen is not

PosturaScreen is a screening and tracking tool. It is not:

• A diagnostic device. We do not diagnose conditions.
• A replacement for radiographic measurement (X-ray, CT, MRI).
• A substitute for clinical examination by a licensed physiotherapist, chiropractor, or physician.
• An exercise prescription tool. We do not recommend specific corrective exercises for individual users.

If you have specific concerns about your posture or any musculoskeletal symptom, consult a licensed healthcare professional.

Frequently asked questions

Does PosturaScreen use AI?

Yes — the detection layer uses YOLO26-pose to locate 17 anatomical keypoints in each photo. But the 17 posture metrics themselves are not AI predictions: each is a deterministic geometric formula (an angle, ratio, or offset) computed over those keypoints. The same keypoints always produce the same metric values, with no per-user black box.

How accurate is 2D photo-based posture measurement?

For repeatability, photographic posture analysis is well-supported: published studies report inter-rater ICCs above 0.97 and test-retest ICCs above 0.77 for photographic postural angles (Hazar et al., 2015), and ICCs around 0.98 for photogrammetric craniovertebral, swayback, and knee measures (Mylonas et al., 2025). For validity against radiographs, a 2025 study of AI-based 2-D posture software found strong correlations with X-ray measures (forward-head vs. craniovertebral angle r = −0.71; hip-knee-ankle r = 0.75) but not identity (Park et al., 2025) — which is why the spinal and angular estimates carry an approx tag.

Can a phone photo replace an X-ray?

No. A photo measures surface landmarks; a radiograph measures bone. The two correlate but are not interchangeable. PosturaScreen is a screening and tracking tool, not a replacement for radiographic measurement.

Why do five metrics carry an "approx" tag?

Forward head, thoracic kyphosis, lumbar lordosis, pelvic tilt, and the Q-angle estimate a 3-D angle or a spinal curve from a flat 2-D photo, where surface contour and body rotation relative to the camera both influence the absolute value. They are reliable for tracking change in the same person over time, but they are not a substitute for radiographic measurement, so they are tagged approx on every report.

How should I take photos so the numbers are reliable?

Use a front and a side photo with the camera at hip height and level (not tilted up or down), about 2-3 meters away, against a plain wall, in fitted clothing, standing relaxed. For tracking, the single most important factor is consistency: re-screen under the same camera height, distance, and lighting, because change over time is what the metrics read most reliably.

Is PosturaScreen a diagnostic device?

No. PosturaScreen does not diagnose conditions. It produces screening measurements for tracking posture over time and for informing a conversation with a licensed physiotherapist, physician, or chiropractor. A diagnosis requires a clinician who can integrate the measurements with a physical examination and history.

Selected references

Peer-reviewed studies on the reliability and validity of photo-based posture measurement that inform this methodology:

• Park, et al. (2025). Validity and Reliability of an Artificial Intelligence-Based Posture Estimation Software for Measuring Cervical and Lower-Limb Alignment Versus Radiographic Imaging. Diagnostics (Basel). PMC12155411
• Mylonas, et al. (2025). Reliability of photogrammetric evaluation of the craniovertebral angle, swayback posture, and knee hyperextension in university students. Journal of Physical Therapy Science. PMC11957747
• Hazar, et al. (2015). Reliability of photographic posture analysis of adolescents. Journal of Physical Therapy Science. PMC4668149

These studies establish two things this methodology relies on: photographic postural angles are repeatable (high inter-rater and test-retest reliability), and AI-based 2-D estimates correlate strongly with radiographic measures without replacing them — the exact basis for reporting useful screening numbers while tagging the radiograph-class metrics approx.

How to reach the methodology team

Researchers, clinicians, or curious readers with feedback, corrections, or questions about our measurement approach: please email support@posturascreen.com. Corrections to this page are tracked per our editorial standards and dated in a changelog at the foot of the article.

Page published: 2026-05-24. Last reviewed: 2026-05-29. Owner: PosturaScreen Editorial Team.