AnyMo Bench: Fine-Grained In-the-Wild Wearable IMU Recognition

Source

• Raw dataset record: source.md
• Normalized metadata: metadata.json
• Additional pinned raw metadata: official_dataset_card.md, official_dataset_info.json, fine150_labels.json, core50_labels.json, and split_subjects.json are preserved in the dataset directory; these auxiliary artifacts are not part of the public wiki artifact allowlist.
• Hugging Face: https://huggingface.co/datasets/CRUISEResearchGroup/AnyMo-Bench
• Introducing paper: AnyMo

Status and credibility

AnyMo Bench is a public Hugging Face dataset released with the 2026 AnyMo arXiv preprint. The pinned repository revision is 2316ef6d3d8544564a55d0cf682aee03d77c2f42, authored on 2026-05-29. The payload is derived from Nymeria, an ECCV 2024 multimodal egocentric daily-motion dataset.

The release includes Parquet shards, explicit label vocabularies, split membership, schema documentation, and build statistics. It is substantially more auditable than an unreleased benchmark, but its headline row count is internally inconsistent and its label pipeline mixes automated and human decisions.

Benchmark contract

AnyMo Bench evaluates activity recognition under a subject-disjoint split and a device shift that preserves body placement:

flowchart LR
  N[Nymeria real IMU] --> Sync[synchronize to 60 Hz]
  Sync --> Windows[variable-length activity windows]
  Windows --> Labels[AnyMo-180 curation]
  Labels --> F[Fine150]
  Labels --> C[Core50]
  F --> US[unseen subject]
  F --> USCD[unseen subject + second co-located IMU]
  C --> US
  C --> USCD

The corpus covers 196 subjects, split into 157 train and 39 test subjects with seed 42. Each row contains 18 numeric features: acceleration and angular velocity from head, left wrist, and right wrist. The pinned export contains 60-300 timesteps per row at 60 Hz.

Four configurations

Configuration	Classes	Subject shift	Device shift
AnyMo-Bench-150-US	150	yes	no
AnyMo-Bench-150-USCD	150	yes	first train unit to second co-located test unit
AnyMo-Bench-50-US	50	yes	no
AnyMo-Bench-50-USCD	50	yes	first train unit to second co-located test unit

The four configurations reuse the same underlying windows. They differ in label taxonomy and device protocol and must not be counted as four independent corpora.

Label provenance

Fine150 is derived from the paper’s AnyMo-180 activity vocabulary after removing classes that are unstable under the subject split or depend too strongly on non-IMU visual/semantic context. Core50 merges Fine150 categories using motion similarity, sample support, and interpretability.

Upstream AnyMo-180 construction combines:

• language-model activity-label proposal;
• normalization and embedding-based semantic clustering;
• expert review of roughly 1,000 candidate labels;
• repeated enum-label classification and language-model adjudication;
• class-wise audits with relabel/drop decisions.

The benchmark therefore has explicit curation provenance but is not a purely human-labeled dataset.

Artifact-grounded statistics

Statistic	Value
Train subjects	157
Test subjects	39
Train rows per configuration	123,874
Test rows per configuration	30,965
Exported rows per configuration	154,839
Paper/card headline windows	154,695
Reported hours	211.6
Train/test Parquet shards per configuration	25 / 7

The 154,839 value is confirmed by official_dataset_info.json. The 154,695 paper/card value is preserved as a conflicting headline claim rather than silently corrected.

Published baselines

The paper trains DeepConvLSTM and MantisV2 for 100 epochs, and adapts COMODO for 20 epochs with MantisV2 and TimeSformer backbones. It reports Acc@1, Acc@5, and macro-F1.

Setting	Strongest Acc@1	Strongest Acc@5	Strongest macro-F1
Fine150 unseen subject	38.5	65.2	22.8
Fine150 unseen subject + cross device	24.0	50.6	8.6
Core50 unseen subject	46.2	78.8	41.3
Core50 unseen subject + cross device	32.6	67.8	23.3

Different models achieve some of these maxima. The table shows benchmark difficulty, not a single-model rank. Cross-device degradation is especially large and should be reported separately from unseen-subject transfer.

Benchmark hygiene

• Pin label space, subject split, device protocol, and dataset revision.
• Keep Acc@1, Acc@5, and macro-F1 separate.
• Do not add row counts across configurations as unique examples.
• Separate raw exported variable-length windows from model-specific interpolation, padding, and splitting.
• Treat the 154,695-vs-154,839 discrepancy as unresolved upstream metadata drift.
• Document whether video is used: COMODO uses synchronized video during training, while the public task surface is IMU recognition.
• Keep paper-reported results separate from independently reproduced results.

License and access

Hugging Face frontmatter declares CC BY-NC 4.0. The card also says that AnyMo Bench follows Nymeria non-commercial research-use terms. These data terms are separate from the AnyMo paper’s CC BY 4.0 license.

The payload is publicly downloadable from Hugging Face. The knowledge base mirrors no Parquet data.

Limitations

• Derived from one upstream collection.
• Only head and bilateral-wrist positions are exposed in the benchmark release.
• Cross-device shift is same-position and co-located, not arbitrary placement transfer.
• Long-tailed labels and mixed automated/human annotation create class and ontology uncertainty.
• The published row count is internally inconsistent.
• No forecasting, action, control-input, intervention, reward, or counterfactual protocol.

Foundation TSFM relevance

Agenda slot	Verdict	Evidence	Missing pieces
Benchmark level	partially closes	Public split, schema, label vocabularies, four explicit transfer configurations, and build metadata.	Resolve row-count drift and add independent reproductions.
Native multivariate structure	partially closes	Eighteen synchronized channels from three body locations.	Fixed low-channel wearable topology; not arbitrary-schema or high-dimensional multivariate data.
Data diversity and long tail	partially closes	196 subjects, 150-class fine-grained label space, and explicit tail-sensitive macro-F1.	Broader populations, devices, placements, and independent source datasets.
Control and counterfactuals	insufficient evidence	None.	Add actions/interventions and state-transition outcomes.

Links into the wiki

• AnyMo Bench
• AnyMo
• AnyMo paper
• Time-Series Classification Foundation Models
• Time-Series Benchmark Hygiene
• Synthetic Data For Time Series

Open questions

• Why do the paper/card headline and artifact split counts differ by 144 windows?
• How much of the cross-device gap comes from calibration, hardware response, synchronization, or remaining orientation mismatch?
• Do class-balanced metrics improve under methods designed explicitly for long-tail wearable recognition?
• How well do models trained here transfer to different manufacturers, placements, sampling rates, and loose mounting?