Danzu

How Does AI Analyse and Score Dancing?

AI scores dancing in three stages: pose estimation converts video into a moving skeleton, audio beat tracking extracts the music's rhythmic grid, and movement features computed from the skeleton are measured against that grid. What comes out are measurements of specific properties — not a judgement of whether the dancing was good.

Updated: 2026-07-20

This page explains the method behind Danzu's scores, including its limits. If you are choosing between dance analysis tools, understanding the pipeline is the fastest way to work out what any of them can legitimately claim.

Stage 1 — Pose estimation: video becomes a skeleton

A pose estimation model processes the video frame by frame and locates anatomical keypoints — shoulders, elbows, wrists, hips, knees, ankles, and so on. Across a clip this produces a time series: the position of every tracked joint at every frame.

Everything downstream depends on this step, which is why filming matters more than people expect. Keypoints get lost when limbs leave frame, when the body is heavily occluded, or when lighting is poor. A clip filmed with feet cropped out is missing the joints that carry most of the information about weight transfer.

Stage 2 — Beat tracking: audio becomes a grid

In parallel, the music is analysed for tempo and beat positions. Beat tracking identifies the times at which beats occur, producing a grid of expected rhythmic moments. This is where a scorer learns what "on time" even means for your particular track — no metronome assumption, no fixed BPM.

Stage 3 — Features: what actually gets measured

The skeleton time series is reduced to movement features. The ones that matter for scoring:

  • Accents. Moments of sharp change in joint velocity — the computational equivalent of a hit. Comparing accent times against the beat grid produces timing.
  • Amplitude and speed. How far and how fast joints travel, which underlies energy.
  • Motion continuity. Whether the body keeps moving rhythmically between accents rather than freezing — the basis of groove.
  • Stability and stop quality. Whether movement resolves cleanly or wobbles, and whether balance holds through transitions, giving control.
  • Variation over time. Whether movement qualities change as the music changes, or stay uniform — which yields expression.

What genre actually changes

Styles genuinely do value different things. Hip hop leans on groove and pocket; K-pop rewards precision; breaking rewards explosive stops. A scoring system can respond to that in two different places, and it is worth being precise about which.

In Danzu, the genre you pick sets how the axes are weighted into your overall score — a hip hop take leans more heavily on groove and timing than a heels take does — and it shapes the coaching you get back. What it does not currently do is change the per-axis measurement itself: timing is computed the same way for every style, so an identical clip scored as hip hop and as contemporary returns the same timing number and a different overall.

Genre-conditioned per-axis thresholds are a harder problem — they need a graded calibration set per style before they mean anything — and claiming them before they exist would be the kind of overclaim this page is meant to argue against. How each style is treated.

Reference-free versus pose comparison

Many dance feedback tools work by comparison: they extract skeletons from your video and a demonstrator's, align them in time, and score the difference. That answers "did you copy this correctly?" and cannot score anything the tool has never seen.

Measuring your movement against the music instead of against another dancer means improvisation becomes scoreable — and means the tool cannot tell you whether you executed a specific routine faithfully, because it never saw the routine. Neither approach is better in general; they answer different questions. More on how the category splits.

What this method cannot measure

Worth being explicit about, because vagueness here is how AI scoring gets oversold:

  • Intent. A deliberate pause and a moment of hesitation look similar to a skeleton.
  • Originality. Novelty requires knowing what has been done before. Feature extraction does not.
  • Cultural fluency.Whether movement is faithful to a style's history and vocabulary is not recoverable from joint positions.
  • Presence and connection. Facial expression, eye focus, and the effect of a performance on a room are outside the measurement.
  • Anything off-camera. A single camera flattens depth — movement toward and away from it is largely lost — and joints out of frame are simply absent.

Repeatability, not truth

There is no objective ground truth for how good a piece of dancing is, so "accuracy" is not quite the right frame. What a scoring system can offer is determinism — the same video producing the same numbers every time — which is what makes two takes months apart comparable.

That is the honest claim: not that a number tells you how good you are, but that a stable instrument applied to comparable clips will show you which direction you are moving. Everything useful about measurement in training follows from repeatability rather than authority.

Common questions

How does AI analyse dance movement?

Pose estimation models locate body joints in each video frame, producing a moving skeleton. Movement features — joint velocity, acceleration, amplitude, and stillness — are computed from that skeleton, then compared against the music's beat grid extracted by audio beat tracking.

How does AI score dancing without a reference video?

By measuring the relationship between your movement and the music rather than the difference between your movement and someone else's. Beat alignment, motion continuity, and dynamic variation are all computable from a single clip, so no demonstrator is needed.

Can AI actually judge whether dancing is good?

It can measure specific, well-defined properties — beat accuracy, movement amplitude, sharpness of stops, rhythmic continuity. It cannot assess intent, originality, cultural fluency, or stage presence. Treat an AI score as a set of measurements, not a verdict.

Is AI dance scoring accurate?

The meaningful question is repeatability rather than accuracy. A well-built scorer is deterministic: the same video always produces the same result, which makes takes comparable over time. There is no objective ground truth for dance quality to be accurate against.

For what the five axes mean in practice, see how Danzu scores dancing. To get a clip scored, see rate my dancing.