Skip to content

Getting the data

IMU-Bench does not redistribute any dataset. Every evaluation set has to be downloaded from its source under that source's licence. This page gives the access steps (checked on 2026-10-14), the licence, the files the harness reads, and the directory layout it expects.

Gates at a glance. No dataset needs a signed data-use agreement or a credentialed account. Three need a free registration, and the rest download directly.

Task(s) Dataset Licence Access
Arm pose, full-body pose, odometry Nymeria CC BY-NC 4.0 email sign-up; download links expire after 14 days
Full-body pose (scorer) SMPL body model SMPL-Model licence (non-commercial) registration
Microgesture OpenWatch CC BY 4.0 Hugging Face login + automatic gate
IMU–text retrieval HumanML3D + AMASS MIT (texts) / AMASS research licence AMASS registration
Parkinson's PADS (PhysioNet) CC BY-NC-SA 4.0 open access, no account
Activity HHAR, MHEALTH (UCI) CC BY 4.0 direct download
Activity PrISM see below partly public
Activity TransfHAR (ours) — release in preparation

Some sets are only partly public, and the TransfHAR release is still in preparation. See Known gaps before planning a full run.

Directory layout

The task cards name our cluster's paths. Put the data under one directory with the layout below and set IMUBENCH_DATA_ROOT to it. Every dataset path in the registry and in the generated configs is then rewritten to that directory.

$IMUBENCH_DATA_ROOT/
  processed_data_body_Ty_rwrist/          Nymeria training recordings, one *_cpfbody_rwrist_imu50hz_wristimu50hz_pose50hz.pkl each (867)
  processed_data_test_body_Ty_rwrist/     Nymeria evaluation recordings (101)
  openwatch_raw/open_watch_without_augmentations/
  HumanML3D/                              HumanML3D texts + UniMTS_data/data/{sim,aug_texts}
  pads_dataset.parquet                    scripts/prepare_pads_dataset.py
  hhar_dataset.parquet                    scripts/prepare_hhar_dataset.py
  mhealth_dataset.parquet                 scripts/prepare_mhealth_dataset.py
  userstudy_watch_dataset.parquet         scripts/prepare_userstudy_watch.py (TransfHAR)
  IMULM_master_dataset_v2.parquet         PrISM, in the shared activity schema

export IMUBENCH_DATA_ROOT=/data/imubench
imu-bench eval --encoder my_encoder.yaml

Two further paths are repo-relative and stay inside the clone: assets/smpl/basicmodel_m.pkl (the SMPL model) and the HF cache for the CLIP text tower used by the retrieval task.

Evaluation datasets

Nymeria (arm pose, full-body pose, odometry)

  • Source: https://www.projectaria.com/datasets/nymeria/, code at https://github.com/facebookresearch/nymeria_dataset. The original release is on the nymeria_dataset_legacy branch; main now holds NymeriaPlus.
  • Licence: CC BY-NC 4.0. Derived data may be shared for non-commercial use, with attribution.
  • Access:
  • Enter your email under "Access the Dataset" on the project page. This accepts the licence and opens the download page.
  • Download the URL file there, or build a filtered one in the explorer (https://explorer.projectaria.com/nymeria).
  • Run download.py -i <urls.json> -o <out> from the Nymeria repository. The links expire after 14 days; fetch a new URL file when they do. The files are personal signed URLs, so never commit them.
  • What the benchmark needs: two data groups per sequence. Skip the head, observer and video groups; the full dataset is about 70 TB.
  • body_motion: body/xdata.npz, the XSens ground truth, about 1.1 TB.
  • recording_rwrist: recording_rwrist/data/motion.vrs and mps/slam/closed_loop_trajectory.csv, about 0.5 TB.
  • Cite: Ma et al., Nymeria: A Massive Collection of Multimodal Egocentric Daily Motion in the Wild, ECCV 2024.

SMPL (full-body pose scorer)

  • Source: https://smpl.is.tue.mpg.de/. Register, download "SMPL for Python", and place the male model at assets/smpl/basicmodel_m.pkl (see assets/smpl/README.md).
  • Licence: SMPL-Model licence, non-commercial scientific research. The model files may not be redistributed.

OpenWatch (microgesture)

  • Source: https://huggingface.co/datasets/pietrobonazzi/openwatch
  • Licence: CC BY 4.0.
  • Access: log in to Hugging Face, accept the gate (approved automatically), then huggingface-cli download --repo-type dataset pietrobonazzi/openwatch.
  • What the benchmark needs: only open_watch_without_augmentations/. In the augmented release, participants 51–85 are mirrored copies of real participants. The CSVs hold raw sensor counts; the task card converts them to m/s² and rad/s.
  • Cite: Bonazzi, Ahmed et al., OpenWatch: A Multimodal Benchmark for Hand Gesture Recognition on Smartwatches, arXiv:2605.04791.

HumanML3D (IMU–text retrieval)

  • Source: https://github.com/EricGuo5513/HumanML3D (code and texts, MIT). The motion comes from AMASS (https://amass.is.tue.mpg.de/), which needs a registration and allows no redistribution of the data or of anything derived from it.
  • Access:
  • Register at AMASS, and download the SMPL+H G archives together with SMPL+H and DMPL.
  • Run the HumanML3D scripts to rebuild the motions, texts and official train/val/test.txt splits.
  • Simulated wrist IMU: the task reads simulated right-wrist IMU and caption paraphrases in the UniMTS format, UniMTS_data/data/sim/<id>.npy and UniMTS_data/data/aug_texts/<id>.txt (UniMTS, https://github.com/xiyuanzh/UniMTS). These two folders are not in the public UniMTS_data.zip on Hugging Face. See Known gaps.
  • Cite: Guo et al., CVPR 2022; Mahmood et al., AMASS, ICCV 2019; Zhang et al., UniMTS, NeurIPS 2024.

PADS (Parkinson's disease)

  • Source: https://physionet.org/content/parkinsons-disease-smartwatch/1.0.0/
  • Licence: CC BY-NC-SA 4.0. Open access: the PhysioNet page reads "Anyone can access the files". No account, training course or data-use agreement is needed.
  • Access: download the ZIP, or wget -r -N -c -np https://physionet.org/files/parkinsons-disease-smartwatch/1.0.0/, then run python scripts/prepare_pads_dataset.py.
  • Cite: Varghese et al., Machine learning in the Parkinson's disease smartwatch (PADS) dataset, npj Parkinson's Disease 10, 9 (2024).

HHAR and MHEALTH (activity)

  • Sources: https://archive.ics.uci.edu/dataset/344 (HHAR) and https://archive.ics.uci.edu/dataset/319 (MHEALTH), both CC BY 4.0, direct download.
  • Build: python scripts/prepare_hhar_dataset.py (reads Watch_accelerometer.csv / Watch_gyroscope.csv) and python scripts/prepare_mhealth_dataset.py.
  • Cite: Stisen et al., SenSys 2015; Banos et al., IWAAL 2014.

PrISM (activity)

  • Source: the PrISM procedure-tracking recordings of Arakawa, Goel et al. (CMU), https://github.com/cmusmashlab/prism. The repository is MIT-licensed; whether that licence covers the data is not stated.
  • What the task uses: 31 participants and 16.3 h over eight procedures. The five classes are our grouping of the procedure steps into activities of daily living (food, household, personal care, appliance, desk/device).
  • Availability: the public download (Dropbox, linked from the repository) holds the cooking and latte_making sessions, about 5.2 of the 16.3 h. See Known gaps.
  • Cite: Arakawa et al., PrISM-Tracker, Proc. ACM IMWUT 6(4), 2022.

TransfHAR (activity)

Sixty activities recorded with a smartwatch for this paper. Its release is in preparation; until then the task cannot be run outside our group.

Pretraining corpus (controlled track only)

Only controlled-track submissions retrain on the IMU-Bench corpus; open-track submissions bring their own weights.

Source Where Licence
WISDM 2019 https://archive.ics.uci.edu/dataset/507 CC BY 4.0
RecoFit https://github.com/microsoft/Exercise-Recognition-from-Wearable-Sensors (git lfs pull) CDLA-Permissive-2.0
PAMAP2 https://archive.ics.uci.edu/dataset/231 CC BY 4.0
SAMoSA https://github.com/cmusmashlab/SAMoSA CMU non-commercial research licence (no redistribution)
Shoaib (complex activities) https://www.utwente.nl/en/eemcs/ps/research/dataset/ no licence stated; cite Shoaib et al., Sensors 2016
UT-Watch https://doi.org/10.18738/T8/NNDFQD CC0 1.0
OCDetect https://zenodo.org/records/13924901 CC BY 4.0
EatingSpeed https://rdr.kuleuven.be/dataset.xhtml?persistentId=doi:10.48804/CN8VBB CC BY-NC-ND 4.0 (no sharing of adapted data)
Motion-X++ https://huggingface.co/datasets/YuhongZhang/Motion-Xplusplus Motion-X licence (CC BY-NC-SA 4.0 + per-subset licences)
BONES-SEED https://huggingface.co/datasets/bones-studio/seed (scripts/download/download_bones_seed.py) BONES-SEED licence (gated; no public re-upload; attribution required)

Caches are built with scripts/build_multidata_cache.py --config configs/base.yaml --model-config <pretrain config> --out-dir <cache dir>. Because of the EatingSpeed, SAMoSA, AMASS and BONES-SEED terms, the built cache cannot be shared as a file.

Known gaps

We are working on each of these. Until they are closed, a third party can run 7 of the 10 tasks end to end: arm pose, full-body pose, odometry, microgesture, Parkinson's, HHAR and MHEALTH.

  • IMU–text retrieval. The simulated wrist IMU and caption paraphrases (UniMTS_data/data/{sim,aug_texts}) are not in the public UniMTS download, and they derive from AMASS, so we cannot redistribute our copy. You can regenerate them with the UniMTS pipeline (pos2bvh.py, then IMUSim bvh2ts.py, then text_aug.py). The paraphrases are written by a language model, so a regenerated set will not match ours exactly.
  • PrISM. Only two of the eight procedures are in the public download.
  • TransfHAR. Release in preparation.
  • Nymeria preprocessing. The converter from the raw Nymeria groups to the per-recording .pkl files is not in this repository yet.