A small bridge improves activity recognition from wrist data
WearableBridge: Task-Guided Latent Transport for Free-Living Activity Recognition · arXiv:2611.09067
In plain words
Wrist sensors record daily movement, but arm gestures and changing device position can make activities hard to tell apart. WearableBridge is a method that learns a small map to adjust wrist measurements before a fixed activity classifier reads them.
It learns from wrist recordings paired with clearer thigh or laboratory recordings, then uses only wrist data for new participants. In PAAWS tests, it improved wrist-only scores in both pairing designs, though it did not match the concurrent thigh score.
The research
Wrist sensors can collect daily movement without a second device, but the signals can blur activities that are clearer in thigh or laboratory recordings.
The result matters for deciding whether a limited set of clearer recordings can help classify wrist data for participants who have no reference recording.
WearableBridge learns a small map from wrist features into a fixed reference classifier, letting wrist-only windows use activity boundaries learned from clearer recordings.
In PAAWS, researchers kept the reference encoder and classifier fixed and compared direct wrist classification with classification after a learned map. MeanFlow improved average activity-level F1 in both reference designs: one paired each wrist window with a simultaneous thigh window, and the other used matched laboratory wrist recordings.
The results show that a small map can make a fixed reference classifier more useful for wrist-only input in these tests. The gains varied by activity and remained below the concurrent thigh score in the Thigh2Wrist comparison.
The main comparison used four activities, current ten-second windows, and one fixed split of 40 PAAWS test participants. The paper also reports that it did not test how gains change as the number of calibration participants shrinks.
What the evidence shows
- A map for wrist signals
WearableBridge is a method for improving activity recognition from wrist accelerometers. It learns a small map that moves wrist features into the activity regions of a fixed classifier trained on clearer reference recordings.
Calibration pairs wrist windows with either simultaneous thigh recordings or matched laboratory wrist recordings; new participants can then be classified from wrist windows alone.
- Mapping helps in PAAWS
In the PAAWS Thigh2Wrist comparison, the same frozen classifier scored mean macro-F1 of 35.07% when applied directly to wrist features and 55.98% after MeanFlow mapping. Macro-F1 averages the F1 score for each activity equally.
This comparison shows that mapping improved wrist-only classification in this four-activity test, while leaving open how close it came to the reference sensor itself.
- The reference still matters
The thigh reference still scored 82.50% macro-F1 when the same classifier read the concurrent thigh windows. That comparison puts the wrist gain in context: WearableBridge made wrist input more useful, but did not make it equivalent to the measured thigh input in this test.
- A second reference design
The laboratory-reference comparison shows the bridge can help when guidance comes from matched wrist recordings rather than a simultaneous thigh sensor. The gain was not uniform across activities, so the average does not imply that every activity improved.
- What reference guidance adds
A PAAWS Thigh2Wrist control compared MeanFlow with a version using the same bridge architecture and activity labels but without its reference-guidance terms. Mean macro-F1 was 55.98% with guidance and 53.21% without it.
This supports a role for those terms in this specific comparison, rather than establishing that they help in every setting.
- Transfer to another cohort
The CAPTURE-24 comparison asks whether a bridge fitted and selected on PAAWS can help in a separate cohort without using that cohort’s labels for fitting or selection. Its improvement supports transfer in this setup, but does not establish performance across other cohorts.
Deep
- Original paper authors
- Jingjing Zou, Yixin Wang
- Submitted to arXiv
- License
- Not specified by source