Key Concepts
Five things determine whether Xianji produces a real measurement or a plausible-looking one. Read these before your first session.
1. Camera viewpoint is not a detail
Section titled “1. Camera viewpoint is not a detail”Footage must be shot from the receiver’s position, looking at the opponent as the athlete would see them in play.
Side-on footage from the stands shows cues a receiver never has access to. The system will still produce a threshold from it, and that threshold will not mean what you think it means. The app warns you on import; it does not stop you.
Results from different viewpoints are not comparable.
2. Clips always end on the contact frame
Section titled “2. Clips always end on the contact frame”Occlusion is achieved by stopping playback early, never by trimming the video file.
This is why the instrument can be accurate: there is no seeking and no frame-hunting during a trial. It also means a clip is never “the shot up to 100 ms before contact” — it is always the full shot, stopped early by a measured amount.
3. A timeout is not a wrong answer
Section titled “3. A timeout is not a wrong answer”If the athlete does not respond before the deadline, it counts as incorrect for the adaptive staircase, but it is stored and reported separately as the no-response rate.
That figure is the commitment measure. An athlete who is 70% accurate but times out on 30% of trials has a different problem from one who is 70% accurate and always commits. Folding them together hides exactly the distinction you need.
4. Chance level differs by sport
Section titled “4. Chance level differs by sport”Badminton has 6 response zones — chance is 16.7%. Tennis has 3 — chance is 33.3%.
Raw accuracy is therefore not comparable across sports. The app always reports guessing-corrected accuracy alongside raw. Use the corrected figure.
5. A non-converged session reports no number
Section titled “5. A non-converged session reports no number”The adaptive staircase needs at least 6 reversals and reasonable consistency between them. If it doesn’t get there, the app says so and reports nothing, rather than emitting a plausible threshold you might act on.
This is deliberate. A refusal is information: it tells you the session didn’t settle, usually because it was too short or performance drifted.
Glossary
Section titled “Glossary”| Term | Meaning |
|---|---|
| Temporal occlusion | Cutting the display at a controlled instant relative to a key event, to determine what information the viewer had used by then |
| Occlusion offset | Milliseconds before contact at which the screen blacks out. Larger = harder = less information given |
| Anticipation threshold | The offset at which the athlete stays ~79% correct. The headline measure. Larger is better |
| Adaptive staircase | 3-down-1-up: three correct in a row makes the next trial harder, one wrong makes it easier. Converges on 79.4% correct without hundreds of trials per level |
| Reversal | A point where the staircase changes direction. The threshold is the mean of the last 6 |
| Standard error (SE) | Precision of the threshold estimate. Treat a change smaller than 2 × SE as noise |
| Chance level | Accuracy from pure guessing. 1/6 badminton, 1/3 tennis |
| Guessing-corrected accuracy | Accuracy with chance removed. The comparable figure |
| No-response rate | Proportion of trials that timed out. The commitment measure |
| Manifest | The manifest.json file describing a footage set: video properties plus every shot’s contact frame, zone, and clip |
| Viewpoint | Where the camera was. Must be receiver |