Two movements,
lined up moment by moment.
Two spikes never take the same time. To compare them fairly, MOTION³ first finds which moment of your rep matches which moment of the reference, then measures the difference at each matched pair.
MOTION AI · Volleyball kinematics
The lab uses the original MediaPipe pose model and retains raw coordinates and visibility. Three-point angles use source-image pixels to preserve the aspect ratio. Shoulder angle is Hip → Shoulder → Elbow; trunk lean is relative to the image vertical. Joints below 0.5 visibility are left unmeasured.
Landmarks use visibility-weighted three-sample smoothing. Angular and landmark velocities use central differences of valid neighboring samples, in deg/video-second and px/video-second. Unknown slow-motion rates cannot establish real motion speeds. Demo events are manually annotated; new events are marked by the user. No calibrated height, ball coordinates or true 3D motion are claimed.
The original STCF-DTW engine remains available for real reference-video comparison. Its parameters are provisional; it does not provide a calibrated volleyball technique grade.
From video to verdict
Why not compare frame by frame?
If you approach a little faster than the reference, a frame-by-frame comparison would set your take-off against the reference's last step and report errors that are only timing. Dynamic time warping (DTW) solves this by stretching and compressing time to find the best match. Classic DTW, though, looks only at joint positions and can match moments that look alike but move differently.
STCF-DTW (spatio-temporal coupled feature DTW) adds motion — velocity, acceleration and coordination between neighbouring joints — to the comparison, respects the body's structure, and limits how sharply timing may bend. The result is an alignment that follows the movement, not just the pose.
The research
The method comes from a peer-reviewed paper on evaluating rehabilitation movements:
Han, Y., Xu, Y., Lu, H., & Gan, Y. (2026). Intelligent Rehabilitation-Action Evaluation System via Enhanced DTW and Deep Learning. In 3D Imaging Technologies (3DIT 2025), Smart Innovation, Systems and Technologies, vol. 500, pp. 255–274. Springer. doi:10.1007/978-3-032-25613-3_22
On the paper's rehabilitation benchmark, STCF-DTW reported the lowest alignment error and the best overlap with expert-marked abnormal segments among the methods compared:
| Method | NDTW ↓ | WD ↓ | IoU ↑ |
|---|---|---|---|
| Traditional DTW | 1.129 | 1.848 | 0.455 |
| Multiscale DTW | 0.746 | 1.406 | 0.631 |
| ShapeDTW | 0.616 | 1.310 | 0.714 |
| DeepAlign | 0.836 | 1.567 | 0.566 |
| STCF-DTW | 0.473 | 1.158 | 0.814 |
NDTW: normalised alignment distance. WD: warping degree. IoU: overlap between detected and expert-marked abnormal segments. Figures as reported in the paper.
What this site adapts
- The paper studies rehabilitation exercises; MOTION³ applies the same alignment to volleyball, fitness and rehab movements.
- Some values the paper leaves open — such as the weights of shape and rhythm, the smoothing window and the severity thresholds — are set for this site and are being reviewed against the original implementation.
- The score is a friendly 0–100 mapping of the alignment distance. It is meant for tracking your own progress, not for comparing athletes.
Limits
- One camera sees the body in two dimensions. Film from the side; movements across the line of sight are measured less reliably.
- Loose clothing, poor light or a busy background can make the body points jump.
- The result compares you with one reference. A different reference, or a different but valid technique, gives a different result.