WristHARBench

Release author: ZipengWu
Version: 1.1.2 (2026-08-31)

WristHARBench is a governed benchmark suite of derived wrist-wearable human activity recognition tasks. It is not a newly recruited cohort or a single newly collected dataset. Every task uses an explicit wrist-stream selection contract and repeated participant-held-out evaluation splits.

WristHARBench v1.1 scope

• 11 public source datasets and 12 task views.
• 9 wrist-ACC main dataset-level entries and 2 auxiliary 9-channel IMU entries.
• 1,214,864 labeled sequences, 310,181,105 multichannel time-step vectors, and 957,295,815 scalar sensor readings.
• 451 task-view participant records and 431 cross-source study-participant records after de-duplicating the paired HARMES sides. The latter is not a claim of globally unique natural persons across independent studies.
• HHAR is excluded from every v1.1 public file and score. Historical HHAR artifacts remain only in the immutable v1 evidence freeze.

The unsuffixed configurations are the public full_task_view release. The frozen manuscript leaderboard is bound to paper_snapshot_20260828. Only Capture-24 differs: 934,762 full windows versus 50,459 snapshot windows. Use capture24_wearable_activity_v1_paper_snapshot_20260828 when comparing with the published frozen scores; do not compare those scores with a model trained on the expanded full Capture-24 configuration.

Evaluation contract

Five repeated participant-held-out assignments use split seeds 1000, 2000, 3000, 4000, and 5000. Despite their column names, they are independent repeated holdouts, not conventional mutually exclusive five-fold cross-validation. Each repeat first holds out approximately 20% of participants for test and then an effective approximately 10% of all participants for validation; integer group rounding applies. Public sequence rows contain fold_0 through fold_4; each value is train, validation, or test. No participant appears in more than one split within a repeat.

Stored signal values are the canonical exporter values before model-time normalization. The formal protocol applies forward-fill then backward-fill imputation within each sequence, followed by per-series z-normalization. timestamp_seconds and signal are fixed-size arrays; signal has shape time × channel in the order given by channel_names.

The formal main track evaluates 10 physical streams × 5 repeated splits × 3 model seeds (42, 43, and 44), for 150 fitted runs. Macro-F1 is computed for each held-out test set. HARMES is one source with paired left/right task views: its score is first computed within each matching split and model seed as (left_macro_f1 + right_macro_f1) / 2. The nine source results have equal weight; mean dataset rank is the primary leaderboard summary. HARMES raw signals are never averaged and its 20 participants are never counted as 40.

One repository, one download

All 12 physical task views, the Capture-24 paper snapshot, catalogs, fold assignments, benchmark code, citations, and source-specific license notices are in Zipeng365/WristHARBench. The repository does not apply one blanket data license: each task view retains the license of its upstream source, as mapped in LICENSE.md and LICENSES/.

Download the complete release in one command:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Zipeng365/WristHARBench",
    repo_type="dataset",
    revision="v1.1.2",
)

Load a catalog:

from datasets import load_dataset

tasks = load_dataset("Zipeng365/WristHARBench", "task_catalog", split="train", revision="v1.1.2")
folds = load_dataset("Zipeng365/WristHARBench", "participant_folds", split="train", revision="v1.1.2")

Configurations

• capture24_wearable_activity_v1 — Capture-24 wrist ACC (934,762 sequences, 151 participant records)
• wisdm_watch_accel_v1 — WISDM watch ACC (74,499 sequences, 51 participant records)
• adl_wrist_accel_v1 — ADL-Wrist right-wrist ACC (4,342 sequences, 16 participant records)
• iuwds_wrist_accel_v1 — IUWDS left-wrist ACC (35,442 sequences, 32 participant records)
• paal_adl_wrist_accel_v1 — PAAL dominant-wrist ACC (14,384 sequences, 52 participant records)
• handy_wrist_accel_v1 — HANDY wrist ACC (7,163 sequences, 30 participant records)
• domino_watch_accel_v1 — DOMINO dominant-wrist smartwatch ACC (11,627 sequences, 25 participant records)
• harmes_left_wrist_accel_v1 — HARMES left-wrist ACC (44,323 sequences, 20 participant records)
• harmes_right_wrist_accel_v1 — HARMES right-wrist ACC (44,323 sequences, 20 participant records)
• gotov_wrist_accel_v1 — GOTOV right-wrist ACC (33,801 sequences, 35 participant records)
• pamap2_hand_imu_v1 — PAMAP2 dominant-wrist 9-channel IMU (7,637 sequences, 9 participant records)
• mhealth_right_lower_arm_v1 — MHEALTH right distal/lower-arm 9-channel IMU (2,561 sequences, 10 participant records)
• capture24_wearable_activity_v1_paper_snapshot_20260828 — exact frozen-paper Capture-24 input (50,459 sequences)

Load one complete task view:

from datasets import load_dataset

ds = load_dataset(
    "Zipeng365/WristHARBench",
    "capture24_wearable_activity_v1",
    split="full",
    revision="v1.1.2",
    streaming=True,
)
row = next(iter(ds))
print(row["series_id"], row["label"], len(row["signal"]), row["channel_names"])

The physical Hugging Face split is named full; the benchmark splits are the five repeated-assignment columns. Do not substitute a random window split. The registered one-command runner and prediction scorer are under benchmark/.

Processing and difference from the upstream sources

Each disclosure below distinguishes the upstream archive from the exact WristHARBench model-input task. Nothing outside the retained scope is silently presented as benchmark input.

Capture-24 — exact retained scope and changes

• Upstream archive: The official release contains approximately 24 hours of 100 Hz Axivity AX3 wrist acceleration for each of 151 participants and is larger than 6.5 GB. More than 200 fine annotation codes are available through the official preparation resources.
• Retained: All 151 participants and all 934,762 official valid 10-second sensor windows are retained in the full task view. Official Walmsley2020 mappings yield four benchmark labels.
• Processing: Form non-overlapping 10-second bins from each participant's first timestamp; require exactly 1,000 source rows and no signal NaNs; take the modal mapped label; retain every fourth sample to convert 100 Hz to 25 Hz and store 250 by 3 arrays. Absolute timestamps and local paths are removed.
• Excluded from model input: Fine-grained annotation codes are coarsened to four labels. Camera images, diary content, absolute time, and non-sensor metadata are not model inputs.
• Exact difference: All eligible official sensor windows are present, but this is a downsampled, label-coarsened derivative rather than a byte-for-byte mirror. The paper snapshot is smaller: 50,459 windows created by an earlier 500-candidate-per-participant cap plus an 80% label-purity rule.

WISDM — exact retained scope and changes

• Upstream archive: WISDM contains 20 Hz accelerometer and gyroscope streams from both a smartphone and a smartwatch for 51 participants performing 18 activities.
• Retained: All 51 official smartwatch-accelerometer participant files and all 18 activity codes are represented, producing 74,499 fixed-length sequences.
• Processing: Parse the official raw/watch/accel files, remove unparsable sensor rows, group samples by activity code, and construct 5-second windows with a 2.5-second stride. Residual fragments shorter than one full window are not emitted.
• Excluded from model input: All smartphone streams and the smartwatch gyroscope are outside this task view.
• Exact difference: Participant and activity scope is complete for the eligible watch-accelerometer stream, but the benchmark is a sensor subset and windowed derivative of the larger phone-plus-watch archive.

ADL-Wrist / WHARF — exact retained scope and changes

• Upstream archive: The source contains a single 32 Hz tri-axial accelerometer on the right wrist for 16 volunteers performing 14 scripted activities. It also ships derived *_MODEL artifact directories.
• Retained: All 16 participants, all 14 activities, and every raw non-MODEL protocol trial are included.
• Processing: Parse numeric trial rows, map official 6-bit values linearly from [0, 63] to [-1.5 g, +1.5 g], and construct 5-second windows with a 2.5-second stride within each trial. Incomplete final fragments are omitted.
• Excluded from model input: The *_MODEL directories are derived templates or model artifacts, not additional participant recordings, and are excluded.
• Exact difference: No participant, activity class, body placement, or raw sensor modality is removed from the source task. The difference is representation: unit conversion and fixed-length windowing rather than raw variable-length trial files.

IUWDS — exact retained scope and changes

• Upstream archive: IUWDS provides synchronized 100 Hz acceleration at the left wrist, left hip, left ankle, and right ankle for 32 adults, plus point labels and participant demographics.
• Retained: All 32 participants are included. The left-wrist axes and all six registered labels are retained, including activity 99 as non-study activity.
• Processing: Select lw_x, lw_y, and lw_z; split the time series at acquisition gaps greater than 0.05 seconds; make 5-second windows with a 2.5-second stride; assign the majority point label; retain only windows with at least 80% label purity.
• Excluded from model input: Hip and ankle axes and demographic variables are excluded from model input. Windows below the 80% label-purity threshold and incomplete tails are omitted.
• Exact difference: The benchmark contains every participant but only one of four placements, and the purity rule makes it a filtered window subset rather than a complete copy of all point-level rows.

PAAL — exact retained scope and changes

• Upstream archive: PAAL v2 publishes 32 Hz Empatica E4 three-axis acceleration for 52 healthy participants performing 24 activities in their natural environments with the device on the dominant hand or wrist.
• Retained: All 52 participants, all 24 activities, and every public acceleration trial are included.
• Processing: Convert released values to g using 64 LSB/g and construct 5-second windows with a 2.5-second stride within each trial. A trial shorter than 5 seconds is end-edge padded to one window; incomplete residual tails of longer trials are omitted.
• Excluded from model input: No participant, activity class, or additional public sensor modality is removed from the PAAL v2 acceleration task.
• Exact difference: The task scope matches the public source acceleration dataset. The benchmark changes only the numeric and temporal representation through unit conversion, fixed windows, and declared short-trial padding.

HANDY — exact retained scope and changes

• Upstream archive: HANDY contains synchronized 52 Hz wrist-worn accelerometer, gyroscope, and magnetometer signals for 30 participants performing ten hand-centered activities.
• Retained: All 30 participant directories and all ten activities are represented. The calibrated low-noise accelerometer axes are the registered input.
• Processing: Read the calibrated timestamp and low-noise acceleration columns, convert m/s² to g by division by 9.80665, repair two documented filename tokens while retaining the raw tokens in provenance, and window within each trial at 5 seconds with a 2.5-second stride.
• Excluded from model input: Gyroscope and magnetometer channels are excluded. Wrist side is recorded as unreported rather than inferred.
• Exact difference: Participant and activity scope is complete, but WristHARBench uses three of the nine motion channels and publishes fixed windows rather than the original trial files.

DOMINO — exact retained scope and changes

• Upstream archive: DOMINO contains smartphone and dominant-wrist smartwatch sensors plus high-level context for 25 users performing 14 activities in indoor and outdoor scenarios.
• Retained: All 25 users and all 14 named activity classes are represented from the LG G Watch R smartwatch acceleration stream.
• Processing: Remove invalid or duplicate timestamps; exclude TRANSITION intervals; split at watch gaps over 1,000 ms; linearly interpolate each remaining labeled chunk to 100 Hz; convert m/s² to g; then make 5-second windows with a 2.5-second stride.
• Excluded from model input: Smartphone signals, smartwatch non-acceleration signals, contextual variables, and TRANSITION intervals are not model inputs.
• Exact difference: This is the complete 14-class participant scope for the selected smartwatch accelerometer, but it is a resampled sensor subset of the larger context-aware multimodal archive.

HARMES — exact retained scope and changes

• Upstream archive: HARMES contains 20 participants, 15 ADLs, left- and right-wrist six-axis motion sensors, environmental sensing, and right-wrist audio. Recordings 01–03 are strongly labeled; recording 04 is free-form.
• Retained: All 20 participants and both wrist acceleration views are retained for strongly labeled recordings 01–03. The target has 15 activities plus the official Null/background class, yielding 44,323 aligned pairs.
• Processing: Drop invalid and duplicate timestamps, resample each wrist acceleration stream to 50 Hz, create non-overlapping 5-second windows on a shared grid, and assign the official maximum-overlap label including Null. Participant IDs and pair keys remain shared across sides.
• Excluded from model input: Both wrist gyroscopes, audio, environmental sensors, and free-form recording 04 are outside the registered task.
• Exact difference: All participants and both wrists are present for the strong-label protocol, but this is an acceleration-only, recordings-01–03 subset of the much larger multimodal archive. Two wrists are paired views, not 40 independent participants.

GOTOV — exact retained scope and changes

• Upstream archive: GOTOV provides activity and energy-expenditure measurements for 35 older adults, including GENEActiv devices at multiple body locations, Equivital, and COSMED signals across 16 everyday activities.
• Retained: All 35 public participant folders and all 16 labeled protocol activities are represented from the right-wrist GENEActiv acceleration file.
• Processing: Keep valid labeled rows; split at label changes, non-positive timestamp steps, unlabelled breaks, or gaps over 100 ms; interpolate each interval to 83 Hz; then construct 5-second windows with a 2.5-second stride.
• Excluded from model input: Unknown or unlabelled rows, ankle and chest sensors, Equivital, COSMED, and energy-expenditure targets are excluded.
• Exact difference: Participant and activity scope is complete for the public right-wrist classification view, but the benchmark is a resampled, acceleration-only subset of the broader multisensor physiology archive.

PAMAP2 — exact retained scope and changes

• Upstream archive: PAMAP2 contains 100 Hz IMUs at the dominant-arm wrist, chest, and dominant-side ankle plus heart rate for nine participants. The release has a 12-activity Protocol set and six optional activities.
• Retained: All nine participants and all 12 Protocol activities are represented. The input is one wrist acceleration range plus wrist gyroscope and magnetometer, nine channels total.
• Processing: Use official Protocol files, remove transient activity ID 0, group retained samples by activity, and make 5-second windows with a 2.5-second stride. Source NaNs are preserved for the registered within-sequence imputation stage.
• Excluded from model input: Six optional activities, heart rate, wrist temperature, the second acceleration range, orientation, and all chest and ankle channels are excluded.
• Exact difference: This is complete for the 12-activity dominant-wrist IMU Protocol task, but it is a placement, channel, and activity subset of the full 18-activity, three-IMU multimodal archive.

MHEALTH — exact retained scope and changes

• Upstream archive: MHEALTH contains 50 Hz motion and vital-sign data for ten volunteers: chest sensors and ECG, a left-ankle sensor, and a nine-axis sensor identified by the column schema as right lower arm. Twelve named activities and null activity ID 0 are present.
• Retained: All ten participants and all 12 named activity classes are represented using source columns 15–23: right-lower-arm acceleration, gyroscope, and magnetometer.
• Processing: Remove activity ID 0, group retained samples by activity ID, and construct 5-second windows with a 2.5-second stride. Channel-dependent upstream units are preserved; incomplete tails are omitted.
• Excluded from model input: Chest acceleration, ECG, left-ankle signals, and null activity ID 0 are excluded.
• Exact difference: Participant and named-activity scope is complete, but the benchmark retains one body placement and 9 of the 23 signal columns before fixed-length windowing.

Scope and limitations

The release supports comparative activity classification under the registered participant-held-out protocol. It does not establish clinical effectiveness, causal effects, universal model superiority, or demographic representativeness. Source archives may contain other sensors or body locations, but those signals are not silently fused into the published task views. PAMAP2 and MHEALTH are auxiliary placement/modality tasks and are not pooled into the 3-axis ACC mean rank.

Sources and attribution

Source	License	Official record
Capture-24	cc-by-4.0	Capture-24: Activity tracker dataset for human activity recognition
WISDM Smartphone and Smartwatch Activity and Biometrics Dataset	cc-by-4.0	WISDM Smartphone and Smartwatch Activity and Biometrics Dataset
ADL-Wrist / WHARF	cc-by-4.0	Dataset for ADL Recognition with Wrist-worn Accelerometer
IUWDS Walk-Climb-Drive	odbl	Labeled raw accelerometry data captured during walking, stair climbing and driving
PAAL ADL Accelerometry	cc-by-4.0	PAAL ADL Accelerometry dataset v2.0
HANDY	cc-by-4.0	HANDY: A Benchmark Dataset for Context-Awareness via Wrist-Worn Motion Sensors
DOMINO	cc0-1.0	Replication data for "DOMINO: A Dataset for Context-Aware Human Activity Recognition using Mobile Devices"
HARMES	cc-by-4.0	HARMES: A Multi-Modal Dataset for Wearable Human Activity Recognition with Motion, Environmental Sensing and Sound
GOTOV	cc-by-nc-sa-4.0	GOTOV Human Physical Activity and Energy Expenditure Dataset on Older Individuals
PAMAP2	cc-by-4.0	PAMAP2 Physical Activity Monitoring
MHEALTH	cc-by-4.0	MHEALTH

Full citation records are in references.bib and the verification ledger is in references_audit.csv. Please cite both WristHARBench and every upstream source whose task view you use.