pakiino/sketch2stl-recognizer

sketch2stl · stroke recogniser (ML1)

Classifies one hand-drawn stroke as line, arc, circle, rect or polyline for the sketch2stl app (draw a 2-D shape → printable STL). Trained from scratch: HistGradientBoostingClassifier over 12 geometric features (sketch2stl.recognizer.features).

Training data

14,608 synthetic hand-drawn strokes (ML1_stroke_recognizer/synthetic_mouse in the dataset): clean Fusion 360 Gallery CAD curves re-drawn by the repo's hand model (sketch2stl.data.synth, style mouse, per-class tremor). Split by CAD model with the official train_test.json; 3,000 held-out test strokes.

Evaluation

test set	n	rules macro-F1	ML1 macro-F1	ML1 accuracy
synthetic, held-out CAD models	3000	0.745	0.964	0.964
real hand-drawn (PK, mouse, never trained on)	325	0.869*	0.765	0.785

Per-class F1 on real strokes: line 0.99 · arc 0.98 · circle 0.79 · rect 0.70 · polyline 0.36

*The rules' thresholds and the hand model's tremor levels were calibrated on these same 325 real strokes, so the real-stroke numbers are optimistic for the rules and somewhat for ML1 too.

Main failure on real strokes: hand-drawn polylines are often read as rectangles.

In the app

The app wraps this model in HybridRecognizer (sketch2stl/recognizer/hybrid.py): ML1 decides line / arc / closed shape, and among closed shapes the least-squares circle / rectangle fits win where they disagree (a circle only if the stroke has at most one sharp corner). On the 325 real strokes that reaches macro-F1 0.91 - an app-quality number, not a held-out score, for the reason above.

Licence

Derived from the Fusion 360 Gallery Dataset (Autodesk) — non-commercial research use only.