PAMAP2¶

Modality: Wearable IMU (accelerometer, gyroscope, magnetometer) + heart rate monitor
Primary Tasks: Daily activity recognition, wearable HAR benchmarking, transfer learning
Scale: 9 subjects (8 train + 1 optional test), 18 labeled activities, 3 IMUs per subject, 100 Hz sampling
License: Creative Commons Attribution-ShareAlike 3.0 (CC BY-SA 3.0)
Access: https://archive.ics.uci.edu/ml/datasets/pamap2+physical+activity+monitoring

Summary¶

PAMAP2 captures high-frequency wearable signals for a range of lifestyle activities including household chores and fitness exercises. It remains a standard benchmark for lightweight models and transfer learning with limited subjects thanks to its synchronized multi-sensor setup.

Reference Paper¶

Attila Reiss and Didier Stricker. "Introducing a New Benchmarked Dataset for Activity Monitoring." ISWC, 2012. PDF

Benchmarks & Baselines¶

DeepConvLSTM - Accuracy: 94.2% (leave-one-subject-out); Ordonez & Roggen, IJCNN 2016.
SelfHAR (self-supervised) - F1: 86.1% (cross-subject); Yuan et al., ICML 2022.
Evaluation typically uses leave-one-subject-out (LOSO) cross-validation; some studies downsample to 33 Hz for efficiency.

Tooling & Ecosystem¶

UCI HAR Utilities for preprocessing and resampling.
tsai library includes dataloaders and benchmark notebooks.
mhealthtools offers feature extraction pipelines for wearable datasets.

Known Challenges¶

Subjects perform free-form activities causing intra-class variability.
Sensor noise increases during high-intensity exercises; filtering and normalization strongly impact performance.
Imbalanced activity durations; consider class-weighted loss or data augmentation.

Cite¶

@inproceedings{reiss2012pamap2,
  title     = {Introducing a New Benchmarked Dataset for Activity Monitoring},
  author    = {Reiss, Attila and Stricker, Didier},
  booktitle = {Proceedings of the 16th International Symposium on Wearable Computers},
  year      = {2012}
}