mp-svm-001: MediaPipe Hand Landmarker + RBF SVM
Reproducible third baseline for 36-class static ASL recognition. It uses the same exact-deduplicated, participant-disjoint split as cnn-001 and mnv4-001.
| Metric | Test result |
|---|---|
| Accuracy | 90.53% |
| Macro precision | 87.04% |
| Macro recall | 90.52% |
| Macro F1 | 88.03% |
| Landmark coverage (test) | 3,589 / 3,589 |
The fitted models/mp_svm_001.joblib is a Python pickle-based artifact: load it only in a trusted environment. Reproduce the entire extraction/training process with notebooks/10_mediapipe_svm_baseline_reproducible.ipynb from the project repository. Dataset, model task asset, upstream split artifact, and their SHA/revision identifiers are frozen in metadata/experiment_config.json.