CAPTURE-24¶
- Modality: Wrist-worn Axivity AX3 accelerometer (100 Hz) + wearable camera + sleep diary
- Primary Tasks: Free-living activity recognition, wearable HAR benchmarking, sedentary behavior analysis
- Scale: 151 participants, 3883 hours total (2562 annotated), 200+ activity labels mapped to Compendium of Physical Activities
- License: Creative Commons Attribution 4.0 (CC BY 4.0)
- Access: https://github.com/OxWearables/capture24
Summary¶
CAPTURE-24 is the largest annotated wrist-worn accelerometer dataset collected in free-living conditions, 2-3 orders of magnitude larger than prior wearable HAR datasets. Participants from the Oxford area wore an Axivity AX3 on the wrist alongside a body-worn camera that provided ground-truth annotations. The combination of scale, ecological validity, and fine-grained labeling (200+ activities) makes it a critical resource for training and evaluating models intended for population-level physical activity epidemiology.
Reference Paper¶
- Matthew Willetts et al. "CAPTURE-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition." Scientific Data, 2024.
PDF
Benchmarks & Baselines¶
- Random Forest (hand-crafted features) - Balanced accuracy ~75% on Willetts split; Willetts et al., 2024.
- SSL-pretrained CNN - Balanced accuracy ~80% (5-class sleep/sedentary/light/moderate/vigorous); Oxford Wearables Group baselines.
- Official evaluation uses per-participant held-out splits; results reported on both 200+ fine-grained and coarser activity groupings.
Tooling & Ecosystem¶
- Official OxWearables/capture24 repository with data loading and preprocessing scripts.
- actipy library for raw accelerometer signal processing and feature extraction.
- ssl-wearables provides self-supervised pre-training recipes compatible with CAPTURE-24.
Known Challenges¶
- Wearable camera annotations can miss activities performed in low-light or private settings; sleep diary supplements these gaps.
- Label granularity varies across participants depending on camera image quality and annotator agreement.
- Class distribution is heavily skewed toward sedentary and sleep; class-weighted evaluation is recommended.
- Raw data is large (~several GB); plan storage accordingly for full 100 Hz signals.