AnyMo Bench

Summary

AnyMo Bench is a public, fine-grained wearable IMU activity-recognition benchmark derived from real Nymeria streams. It evaluates unseen-subject generalization and same-position cross-device transfer using synchronized head, left-wrist, and right-wrist inertial measurements.

Dataset contract

FieldValue
SourceNymeria real in-the-wild wearable IMU
Subjects196: 157 train, 39 test
Body positionshead, left wrist, right wrist
Sampling60 Hz
Numeric features18 acceleration/gyroscope channels
Window length60-300 timesteps in the exported release
Label spacesFine150 and Core50
Protocolsunseen subject; unseen subject + cross device
Public artifactHugging Face Parquet release
License surfaceCC BY-NC 4.0 frontmatter plus Nymeria non-commercial terms

Configurations

  • AnyMo-Bench-150-US: Fine150, first IMU for train and test.
  • AnyMo-Bench-150-USCD: Fine150, first IMU for train and second co-located IMU for test.
  • AnyMo-Bench-50-US: Core50, first IMU for train and test.
  • AnyMo-Bench-50-USCD: Core50, first IMU for train and second co-located IMU for test.

Each configuration has 123,874 train rows and 30,965 test rows in the pinned artifact metadata.

Why it matters

AnyMo Bench isolates a practically important setup shift: the body location is held fixed while the physical IMU unit changes. The reported results show a large drop from unseen-subject to cross-device classification, so device invariance should be measured separately from subject invariance.

The benchmark also provides a stronger long-tail test than typical small-label wearable HAR suites: Fine150 keeps 150 activity classes, while Core50 merges them into a coarser but still substantial vocabulary.

Row-count tension

The paper and data card state 154,695 windows. Their split counts, and the pinned dataset_info.json, instead contain 123,874 train plus 30,965 test rows, totaling 154,839 per configuration. The wiki treats 154,839 as the exported artifact count while preserving 154,695 as the paper/card claim.

Benchmark hygiene

  • Do not aggregate all four configurations as independent samples; they reuse the same window inventory.
  • Report label space, subject protocol, and device protocol with every score.
  • Keep Acc@1, Acc@5, and macro-F1 separate; long-tail performance is not captured by top-k accuracy alone.
  • Pin the Hugging Face revision and label vocabularies.
  • Distinguish exported variable-length windows from model-specific padding, splitting, or interpolation.
  • Keep automated label construction, human review, and downstream evaluation as separate provenance layers.

Caveats

  • One upstream dataset and three body positions.
  • Cross-device evaluation uses co-located devices and does not cover arbitrary new placement.
  • Label construction includes language-model proposal and adjudication.
  • Fine150 and Core50 remain long-tailed.
  • No forecasting, action, intervention, reward, or counterfactual interface.