YiquanLi/IMU-Bench-results

IMU-Bench: results and leaderboard data

The per-cell result files, task cards, encoder cards and splits behind IMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks (Li, Bradshaw, Gao, Ahuja, 2026; arXiv:2610.19481).

• Website and leaderboard: https://spicexlab.github.io/imubench/ (also served as a Space: https://huggingface.co/spaces/YiquanLi/IMU-Bench)
• Code and harness: https://github.com/SPICExLAB/IMU-Bench
• Reference model: https://huggingface.co/YiquanLi/InrtlWrist

The benchmark's evaluation datasets (Nymeria, OpenWatch, HumanML3D, PADS, HHAR, MHEALTH, PRISM, TransfHAR) are not redistributed here; the code repository's README lists their sources. This repository holds what the benchmark produces and what defines it.

Contents

results_long.csv                                      every cell as one row (task, encoder, protocol, seed, primary metric and value, all metrics as JSON); what the dataset viewer shows
results/<task>/<encoder>__<protocol>__s<seed>.json   one file per scored cell (2,300+ cells, every task incl. ablation variants)
registry/tasks.yaml                                   task cards: dataset, split rule, input window, decoder, recipe, selection, metrics, scorer
registry/encoders.yaml                                encoder cards: run / checkpoint, architecture flags, pretraining corpus
registry/protocols.yaml                               probe / finetune / scratch
registry/splits/                                      fixed split files that are not derived from a rule (Nymeria participant thirds)
leaderboard.json, LEADERBOARD.md                      the leaderboard generated from results/ by `python -m bench.leaderboard`
paper_results/tab2_main_results/                      tidy CSVs behind Table 2, the radar figure and the per-task appendix tables

A result file records the task, encoder, protocol and seed, every metric of the cell, the evaluation split, the run and the selected checkpoint step, the selection rule, the job ids and the provenance of the run:

{"task": "microgesture_openwatch", "encoder": "mae_5src", "protocol": "probe", "seed": 42,
 "metrics": {"clip_macro_f1": 0.647, "clip_acc": 0.678, "window_macro_f1": 0.494, "window_acc": 0.536},
 "primary_metric": "clip_macro_f1", "eval_split": "test", "run": "...", "checkpoint": "...", "checkpoint_step": 1200,
 "selection": {...}, "job_ids": [...], "provenance": {...}, "status": "ok", "notes": "", "schema_version": 1}

The paper's ten tasks are arm_pose_nymeria, full_body_pose_nymeria, odometry_nymeria_local, microgesture_openwatch, language_humanml3d, pd_pads, har_hhar, har_mhealth, har_prism and har_userwatch (TransfHAR); the six matched encoders are mae_5src (InrtlWrist), limubert_5src, simmtm_5src, focal_5src, relcon_5src and sslwear_5src. Other task ids are protocol ablations (decoder ladders, rotation variants, release splits) and other encoder ids are ablations, controls and external systems; the cards say which.

Leaderboard tracks

leaderboard_tracks.json holds the two tracks shown on the website, generated from results/ by imu-bench leaderboard. Both use the same ten task cards and scorers.

• Controlled track: pretraining corpus (197,094 windows) and backbone (21 M ViT) are fixed, and only the objective differs. It compares objectives.
• Open track: any architecture, size, corpus and objective. Weights must be public, and the pretraining corpus must not contain evaluation recordings. Released models from other groups are run here from their public checkpoints, and their authors are welcome to resubmit.

Both tracks print two decoder-alone controls in every column: a never-trained frozen encoder and the raw input with zero layers. Each row also says who ran it.

Submitting to the leaderboard

git clone https://github.com/SPICExLAB/IMU-Bench && cd IMU-Bench && pip install -e .
export IMUBENCH_DATA_ROOT=/data/imubench            # DATASETS.md in the code repository
imu-bench eval --encoder my_encoder.yaml            # any PyTorch model via the plugin interface; all 20 cells
imu-bench check my_encoder                          # complete, recipe identical to the task cards, card complete
imu-bench package my_encoder --out submission/      # results/, registry/<id>.yaml, configs/, check.json

Then open a pull request to this dataset:

from huggingface_hub import HfApi
HfApi().upload_folder(repo_id="YiquanLi/IMU-Bench-results", repo_type="dataset", folder_path="submission",
                      create_pr=True, commit_message="Add my_encoder")

We merge after re-running imu-bench check on the files and regenerate the leaderboard. A cell counts only if it comes from the released task cards: same raw input (one wrist, six channels, 50 Hz, the released windows and splits), decoder, recipe, split, selection rule and scorer. Input normalisation and other preprocessing inside the encoder are the submitter's choice, declared in the card (normalize_method). Systems trained under their own task recipes are listed for orientation and are not ranked.

Citation

@article{li2026imubench,
  title   = {IMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks},
  author  = {Li, Yiquan and Bradshaw, Aidan and Gao, Chenfeng and Ahuja, Karan},
  journal = {arXiv preprint arXiv:2610.19481},
  year    = {2026}
}