ActiNet: an open-source tool for activity intensity classification of wrist-worn accelerometry using self-supervised deep learning
Aidan Acquah, Shing Chan, Aiden Doherty
Summary
ActiNet is the cleanest public recipe for "SSL wrist-ACC encoder → fine-tune → HMM" for 24-h activity and sleep composition. That maps directly onto Clarity V2's accelerometer branch and gives a strong baseline for day-level activity summaries.
Key facts
- Form Factor
- wrist
- License
- Oxford Academic Use Licence (non-commercial; commercial use needs a separate Oxford licence)
Reported benchmarks
Human activity recognition (IMU)
capture-24 · Cohen's κ ↑
same 4-class intensity protocol as above; RF+HMM baseline 0.800
Human activity recognition (IMU)
capture-24 · Macro-F1 ↑
Sleep/sedentary/light/MVPA (Walmsley), 30-s @30 Hz wrist ACC, 151 pts, nested 5-fold group CV; κ 0.860; RF+HMM 0.764.
Human activity recognition (IMU)
RealWorld HAR · Macro-F1 ↑
Deployed CAPTURE-24 model, wrist ACC; RF+HMM 0.531.
Human activity recognition (IMU)
WISDM AR v1.1 · Macro-F1 ↑
Deployed CAPTURE-24 model, labels mapped to intensity; RF+HMM 0.679.
Updated: Oct 13, 2026
Related Research
- Human activity recognition (IMU) leaderboard
Benchmark results for human activity recognition (imu)
- An open-source deep learning model for classifying real-world activity intensities in persons with chronic pain using a wrist-worn accelerometer
A rare CC BY 4.0 wrist-ACC activity-intensity model with weights, so unlike ActiNet (Oxford academic
- AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild
Treats wearable setup variation as structured body-surface geometry: dense physics-grounded IMU simu
- BayaHAR: Lightweight Bayesian Few-Shot User Adaptation for On-Device Personalized Human Activity Recognition
Turns pretrained HAR classifiers into prototypical networks with prior prototypes that keep zero-sho