Results¶

The primary analysis evaluates complete activity records on an independent external test set.

37test recordings

114reference segments

0.90micro-F1

Evaluation protocol¶

Predictions are matched one-to-one with same-class reference segments at IoU > 0.5. This record-level protocol penalizes missed and false activities, fragmentation, merging, wrong classes, and shifted boundaries.

Data role	Recordings	Use
Training	80	Model fitting and training-stage checkpoint selection
Development / calibration	20	Temporal-policy calibration and diagnostics
Independent external test	37	One final evaluation after all choices were frozen

External labels were not used to select checkpoints, fusion rules, TRL parameters, or reported variants. Mean-user F1 weights each user equally; micro-F1 pools TP, FP, and FN over the complete set.

External test¶

Fixed variant	Mean-user F1	95% CI	Micro-F1	TP / FP / FN
5 s + 8 s + TRL	0.88	0.80–0.94	0.88	98 / 11 / 16
Three-model average + TRL	0.88	0.80–0.94	0.89	98 / 9 / 16
Three-model weighted + TRL	0.89	0.81–0.94	0.89	99 / 9 / 15
LBSA + relaxed Top-K	0.88	0.80–0.95	0.88	103 / 17 / 11
LBSA + TRL	0.89	0.82–0.94	0.90	99 / 7 / 15

Primary result. LBSA + TRL retains the top rounded F1 while producing the fewest false-positive records among the compared variants.

By activity¶

Activity	Reference segments	TP / FP / FN	Precision	Recall	F1	Matched IoU
Badminton	32	26 / 3 / 6	0.90	0.81	0.85	0.87
Rope skipping	20	19 / 0 / 1	1.00	0.95	0.97	0.84
Dumbbell fly	20	19 / 1 / 1	0.95	0.95	0.95	0.78
Running	20	18 / 1 / 2	0.95	0.90	0.92	0.82
Table tennis	22	17 / 2 / 5	0.90	0.77	0.83	0.86

Rope skipping has the highest F1. Table tennis has the lowest recall, indicating that weak or ambiguous evidence still causes missed records.

Representative cases¶

Ground truth and fixed temporal fusion variants for one successful and one partially failed activity timeline — One successful case and one partial failure.

Scale arbitration improves a boundary in the successful case. The failure case retains a miss and a false alarm where posterior evidence is weak; TRL cannot reconstruct activity evidence that the classifiers never provide.

Limitations¶

Evaluated scope

Results measure segment-record quality, not clinical, coaching, or safety value.
Evidence covers five activities under the studied device, placement, and protocol.
New devices, placements, users, activities, and deployment conditions require new validation.
Dense sessions and adjacent same-class events remain difficult.

Development diagnostics, public-dataset portability checks, and Android implementation evidence are reported separately in Supplementary analyses. Reproduction commands and asset boundaries are collected on the Reproduce page.