Analysis, accuracy & evidence FAQ

1. Tracking — body landmarks are located in a sample of frames, in your browser.
2. Phases — the motion is split into its phases (for example setup, top, impact and finish for a golf swing).
3. Reads — rotation, posture, sequencing, balance and tempo are estimated, each with its own confidence.
4. Quality — the capture is graded, and reads the footage cannot support are lowered or withheld.
5. Decision — rules written for the sport rank what they found and choose one priority to work on.

The result does not depend on a language model deciding what is wrong with your swing. Optional AI features sit alongside it: AI coaching, where available, explains findings that were already worked out this way, and an AI second opinion you can choose to run gives a separate read from sampled still frames. Neither one picks your One Fix.

Phone video from one camera gives useful training estimates, not lab measurements. Accuracy drops with a poor angle, low light, motion blur or a cropped body, and the quality check lowers confidence when that happens.

Validation is deepest for golf. For the other live sports, reads are directional and their confidence is capped lower. Per-sport detail is on the video analysis methodology page.

Do not treat one analysis as a verdict. Look for a finding that repeats across clips, and use a retest to check a change.

• Measured — for example ball speed or club path from a launch monitor you imported. 3D is only called measured with a verified camera calibration, which an ordinary phone recording — even from two angles — does not have.
• Imported — values that came in from another device keep that device's measurement; SwingVantage does not upgrade or change them.
• Estimated — body angles, rotation and tempo from phone video: useful for training, not lab-grade.
• Inferred — an interpretation from indirect evidence, such as reading hand path as a proxy for something the camera cannot see.
• Self-reported — something you told SwingVantage, such as your usual miss.
• Not measured — something the current evidence cannot support, which is left out rather than guessed.

The labels are there so a number never looks more certain than the evidence behind it.

Confidence is worked out per read, not per video. A clip can support a confident tempo read and only a low-confidence posture read at the same time, for example because the camera was high.

Confidence goes down for capture problems (dark, blurred, cropped or shaky footage), for reads the camera angle cannot see well, and for sports whose analysis is less validated. Clean footage from an angle that shows the movement well is what raises it.

A clean clip of an unusual swing can produce a confident read of that unusual swing. That is why SwingVantage asks you to retest under the same conditions and looks for patterns that repeat before calling something a habit.

Confidence is also not a score of how good your swing is — a low-confidence read can come from a great swing filmed badly.

Current ceilings: Golf 100%, Baseball 90%, Slow-Pitch Softball 90%, Fast-Pitch Softball 90%, Tennis 85%, Pickleball 80% and Padel 80%. These are the same limits the analysis applies to your numbers, and they are published per sport on the video analysis methodology page.

For golf the analysis supports pose landmarks, phase timing, club/path proxies, putting read and golf-specific diagnosis. What each other sport reads — and what it deliberately does not claim — is listed on its own FAQ page and on the video analysis methodology page.

Directional reads are still useful: they point you at the biggest movement pattern to change, and a retest shows whether it moved.

Fast moments also fall between frames. At 30 frames per second the instant of contact in a fast swing is usually not captured exactly, so the strike frame and tempo are estimated.

This is why camera angle matters so much, why some reads are labelled inferred, and why the quality check names which reads a clip limits. A second camera angle improves depth; a verified calibration is needed before anything is called measured.

Camera height matters too: a phone well above or below chest height tilts the apparent angle of your spine and turn, and the quality check flags that.

For comparisons, keep the angle, height and distance the same every time — otherwise a change in the result may be a change in the camera, not in you.

Check the capture grades of both clips first. If either was Usable or lower, the difference may be the footage. If both were Good or Excellent and filmed the same way, the difference is more likely real swing-to-swing variation.

Several swings in one clip help here: Motion Lab reports your most repeated miss and how consistent you were, which is a steadier guide than any single rep.

If it could not find a body at all, no motion reads are shown. If it found one only partly, the clip is graded not defensible, whatever it could read is shown with heavily capped confidence, and the result tells you to re-film before relying on it.

When the read was very poor, you may also be offered an optional cloud re-analysis, which only runs if you explicitly agree.

Estimates are most reliable when the camera looks straight at the movement being read and the whole body is in frame. Angles that point toward or away from the lens are the hardest to estimate from one camera.

Use them the way a coach uses a video: to see direction and change over time, not to chase a single decimal point.

For some swings — such as a baseball or softball hitting swing, or a racquet stroke — the results may show a rough bat or racquet speed such as "Bat ~68 mph". It is labelled as an estimate: it comes from a path partly projected from your arm movement, scaled by your body size, and a single camera under-reads the true peak, so it is not a substitute for a launch monitor or bat sensor. Golf club speed is not shown from video, and exit velocity and pitch speed are not read from video at all.

Where the analysis describes what the ball is likely to do, it is an interpretation of your movement, labelled as such — not a flight measurement.

That means every finding can be traced back to specific evidence in your clip, with its confidence and label. Where optional AI coaching features are available, they explain results that were already worked out; they do not decide what your fix is.

With a free account you can also choose to run an AI second opinion on a clip you have just analysed. It sends a small sample of still frames — never the video file — to a third-party AI service for a separate read, runs only where that review is available, and does not replace your One Fix. See what leaves your device.

Motion Lab summarises several swings in one clip (most repeated miss, consistency), and a retest checks whether a finding is still there after practice. Both are ways of asking "does this repeat?" before you spend weeks on it.

Some of these can come in from a device you import from, and they keep that device's label. The rest are simply not claimed — the product leaves them out rather than guessing.

Each sport's page lists what its analysis reads and what it does not claim, taken from the same list the analysis itself uses.