AITRAF Demo

FUTURE IS NOW. "aitraf-project" - AI models that predicts & evaluates aggressive inline tricks.

» DEMO APP «

The inference in this demo is simulated: nothing actually runs when you press run, and the results are pre-computed offline and replayed here to avoid the high model compute costs.

The predictions themselves are real: the trick, stars and confidence you see are what the model actually returned for that clip.

» DATASET & LIMITATIONS «

The models were trained on approx 1k short clips, each of a single grind trick, on the same obstacle from the same camera position. They were not generalized to work on every obstacle or environment.

7 grind classes were trained:

soul
top-soul
mizou
ao-soul
fs-royale
bs-royale
fs-savanah

3 quality grades were labeled:

★Failed: fell, landed on top of the obstacle (darkside), or landed on one leg, etc.
★★Average: landed, but with visible mistakes, usually grinding only small part of the obstacle.
★★★Clean: grinded the whole obstacle with no obvious mistakes.

» RESULTS «

Models are tested on data unseen by the training set and compared against baseline/dummy predictions. Arrows indicate whether higher or lower values are better.

Trick classification
Model	Accuracy ↑	Macro F1 ↑
Dummy model (majority class)	0.160	0.039
aitraf_trick_class_V1	0.840	0.838

Trick score
Model	AMAE ↓	QWK ↑
Dummy model (median class)	0.667	0.000
aitraf_trick_score_V1	0.218	0.775

What these metrics mean

Accuracy is how often the model named the trick correctly.

Macro F1 (macro-averaged F1 score) is the same idea, except every trick counts equally, so the model cannot look good just by getting the common tricks right.

AMAE (average mean absolute error) is how far the model's star rating lands from the real one, averaged over every clip.

QWK (quadratic weighted kappa) is how closely its ratings line up with the real ones, counting a two-star miss as much worse than a one-star one.

» CREDITS «

Author: Henrikas Antanas Girdzijauskas (@h3en1x)
Supervisor: Prof. dr. Aistis Raudys
Reviewer: Prof. dr. Olga Kurasova

I want to thank close friends at h2bc (@_h2bc) and the local skate community at Vilnius Rolling (@vilniusrolling) for helping to film the training dataset.