WEAR¶
- Modality: 4x Bangle.js smartwatches (accelerometer + gyroscope, 50 Hz) + GoPro egocentric video
- Primary Tasks: Outdoor sports activity recognition, wearable HAR benchmarking, egocentric activity verification
- Scale: 22 participants, 18 workout activities, 11 outdoor locations, ~18 hours of data
- License: Creative Commons Attribution 4.0 (CC BY 4.0)
- Access: https://github.com/mariusbock/wear
Summary¶
WEAR is an outdoor sports dataset combining data from four open-source Bangle.js smartwatches (both wrists and ankles) with synchronized egocentric GoPro video. The focus on outdoor workout activities across diverse real-world locations distinguishes it from lab-based wearable datasets. The use of affordable, open-source hardware (Bangle.js) lowers the barrier for replication and extension. WEAR supports research in multi-device wearable fusion, egocentric video verification of IMU-based predictions, and robust HAR under uncontrolled outdoor conditions.
Reference Paper¶
- Marius Bock et al. "WEAR: An Outdoor Sports Dataset for Wearable and Egocentric Activity Recognition." IMWUT, 2024.
PDF
Benchmarks & Baselines¶
- DeepConvLSTM - F1-score reported on LOSO evaluation; Bock et al., IMWUT 2024.
- Transformer baseline - Competitive F1 with attention-based temporal modeling on inertial streams.
- Video baseline - Egocentric video recognition accuracy provided for comparison with wearable-only models.
- Official evaluation protocol uses leave-one-subject-out (LOSO) cross-validation.
Tooling & Ecosystem¶
- Official mariusbock/wear repository with data loaders, preprocessing scripts, and baseline implementations.
- Built on Bangle.js open-source smartwatch platform; firmware and data collection apps are publicly available.
- Compatible with PyTorch and standard wearable HAR pipelines; sliding window segmentation scripts included.
Known Challenges¶
- Outdoor recording introduces GPS drift, variable lighting for video, and environmental noise in accelerometer signals (e.g., wind, uneven terrain).
- Bangle.js consumer-grade sensors have higher noise floors compared to research-grade IMUs; calibration may improve results.
- Activity transitions and rest periods between exercises require careful segmentation.
- 22 participants is moderate; cross-dataset evaluation recommended for generalization claims.
Cite¶
@article{bock2024wear,
title = {WEAR: An Outdoor Sports Dataset for Wearable and Egocentric Activity Recognition},
author = {Bock, Marius and Malmgren-Hansen, David and Moeslund, Thomas B. and others},
journal = {Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies},
year = {2024}
}