154,695 eligible activity windows from 196 subjects, covering 211.6 hours of real in-the-wild IMU data. The benchmark is built from real wearable IMU streams in the Nymeria dataset and provides unseen-subject and cross-device evaluation settings for wearable motion recognition. It was created by the CRUISEResearchGroup and last updated in May 2026.
Use Cases
- Training models for fine-grained activity recognition based on synchronized IMU streams.
- Evaluating model generalization in unseen-subject and cross-device settings as described in the benchmark.
- Benchmarking HAR algorithms on real-world, in-the-wild motion data.
Strengths
- Contains 154,695 activity windows from 196 subjects, providing a substantial sample size.
- Covers 211.6 hours of real in-the-wild IMU data, representing realistic usage scenarios.
- IMU streams are synchronized to a common 60 Hz temporal grid, ensuring temporal alignment.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
Provenance
- Source
- CRUISEResearchGroup via Hugging Face.
- Collection Method
- Built from real wearable IMU streams in the Nymeria dataset.
- Freshness
- Last updated 2026-05-28 23:38:14; freshness should be verified.