bubblesmal/spotter-models

Spotter models

Models for Spotter, a quiet AI spotter for new gym-goers and people returning from injury (Knight Hacks IX). The iPhone app runs the pose model on-device; the form model scores reps. The app reads manifest.json on launch and downloads any model that changed.

Manifest version 2, published 2026-10-10T17:03:35+00:00.

Model	Format	Path	Size
form-quality-label-onnx	onnx	form/form_quality_label.onnx	0.9 MB
form-quality-score-onnx	onnx	form/form_quality_score.onnx	0.2 MB
pose-256-coreml	coreml	pose/yolov8n-pose-256.mlpackage	6.7 MB
pose-320-coreml	coreml	pose/yolov8n-pose-320.mlpackage	6.7 MB
pose-320-onnx	onnx	pose/yolov8n-pose-320.onnx	13.3 MB

Pose: YOLOv8n-pose

• Ultralytics YOLOv8n-pose (COCO keypoints), exported at 320 and 256 px to Core ML (FP16, no NMS) and at 320 px to ONNX.
• Output: one tensor [1, 56, N] of raw candidates: 4 box values, person confidence, then 17 joints x (x, y, visibility), with sigmoid already applied. Decoding and orientation rules: docs/pose/how-it-works.md and docs/pose/pose-result-contract.md in the Spotter repo.
• Video never leaves the phone; only joint positions are sent to the cloud.

Form quality

• spotter.ml.form_quality version 4 from Databricks Unity Catalog, as two ONNX graphs: label (clean, partial_range, knees_in, forward_lean, fast_lowering, other) and score (quality, clip to 0-1).
• Inputs, one row per rep: exercise_id, range_deg, weight_kg. exercise_id is a string; unseen exercises are ignored. Numeric inputs may be missing (NaN).
• Held-out metrics: label_accuracy 0.188, label_macro_f1 0.2427, quality_mae 0.0807, quality_r2 0.153, test_lifters 40.0, test_reps 35816.0.
• Limitations: a baseline gradient-boosted model trained on labels derived from the phone's rule flags and synthetic sessions, not hand-labelled clips. Expect it to be replaced.

Intended use

Real-time rep counting and form feedback for recreational lifting. Not a medical device and not for diagnosing injuries.

License

The pose model is derived from Ultralytics YOLOv8 and is released under AGPL-3.0, the license of the original weights.