Dataset & Model Training
Build a real wakeboarding trick dataset
Every training example here comes from a clip you uploaded: the pose track measured in your browser, the motion features derived from it, and the trick name you confirmed yourself. Nothing is simulated. Until there is enough real labelled data, this page stays in data-collection mode and no accuracy is reported.
Labelled examples per class
MVP vocabulary: frontside and backside are counted together as “180” for now, and 360s count as Other. Saved labels are never rewritten — the grouping is applied when the data is read, so the vocabulary can widen later without losing anything.
Readiness
Counting saved records for —
- Usable examples
- 0 / 15
- Runs labelled
- 0 / 2
- Classes with enough data
- 0 / 2
- Feature dimensions
- 0
Data-collection mode: not enough labelled examples to train a useful model yet. Keep uploading clips and confirming the trick on each detected move.
No trained model yet
There is no trained classifier in this project, so no accuracy, confusion matrix or learned prediction can be shown. Analyses currently name moves with explicit, inspectable kinematic thresholds — that is stated on every result. To reach a first trained model you still need: at least 5 confirmed examples in each of at least 2 classes, 15 examples in total, drawn from at least 2 different runs. The fastest way there is to label every detected move in each run you upload, not just one. Clips filmed from a steady side-on angle, with the whole rider in frame, give usable pose features.
Recent labelled examples
No labelled examples yet. Upload a run, then confirm the trick on each detected move.