Direction Prediction on Tennis Stroke Dataset 11 videos
61.9AccuracyTennisTransformerGPU
Evaluation Results
| Method | Links | |
|---|---|---|
| TennisTransformerGPUPose tool=MediaPipe Pose Landmarker, Landmarks=33 (image + world coords.), Coordinate space=World (metric, hip-origin), Depth / z-axis=Yes, Visibility scores=Yes, Stroke segmentation=Automated velocity-based, Player-adaptive thresholds=Yes, Model type=Transformer Encoder, Sequence attention=Yes (4 heads, 4 layers), Multi-task learning=Yes (2 learned + feedback), Hardware=Kaggle free-tier (T4), Evaluation scope=full-stroke2026.06 | 61.9 |