CRUISEResearchGroup/AnyMo-Bench

AnyMo Bench

AnyMo Bench is a challenging fine-grained in-the-wild HAR benchmark built from real wearable IMU streams in the Nymeria dataset. It provides unseen-subject and cross-device evaluation settings for wearable motion recognition.

For general project information, see the AnyMo project page. For more technical details, see the AnyMo paper. The code is available at Breezelled/AnyMo.

The benchmark contains 154,695 eligible activity windows from 196 subjects, covering 211.6 hours of real in-the-wild IMU data. IMU streams are synchronized to a common 60 Hz temporal grid. Each row contains one activity window with an imu array of shape [T, 18], where T <= 300 and the 18 channels concatenate one selected IMU from each of Head, Left Wrist, and Right Wrist.

Configurations

• AnyMo-Bench-150-US: Fine150 label space, unseen-subject evaluation, train/test use the first co-located IMU at each body position.
• AnyMo-Bench-150-USCD: Fine150 label space, unseen-subject + cross-device evaluation, train uses the first co-located IMU and test uses the second co-located IMU at each body position.
• AnyMo-Bench-50-US: Core50 label space, unseen-subject evaluation.
• AnyMo-Bench-50-USCD: Core50 label space, unseen-subject + cross-device evaluation.

from datasets import load_dataset

dataset = load_dataset("CRUISEResearchGroup/AnyMo-Bench", "AnyMo-Bench-150-US")

Baseline Results

The table below reports IMU-based HAR baseline results from the AnyMo paper. Metrics are percentages. The best result in each setting/metric is shown in bold, and the second-best result is underlined.

Model	Acc@1	Acc@5	Macro-F1
Fine150 / Unseen Subject
DeepConvLSTM	35.3	63.0	17.2
MantisV2	38.5	65.2	22.8
COMODO	37.8	65.2	16.0
Fine150 / Unseen Subject + Cross Device
DeepConvLSTM	1.9	9.5	0.4
MantisV2	14.4	39.7	8.6
COMODO	24.0	50.6	8.0
Core50 / Unseen Subject
DeepConvLSTM	43.2	75.4	34.5
MantisV2	45.8	76.8	41.3
COMODO	46.2	78.8	37.3
Core50 / Unseen Subject + Cross Device
DeepConvLSTM	1.8	12.1	0.6
MantisV2	16.6	48.7	18.4
COMODO	32.6	67.8	23.3

Schema

• sample_id: unique row identifier within the exported benchmark.
• session_id: Nymeria sequence/session directory name.
• subject_id: Nymeria anonymized subject identifier.
• body_position: concatenated body-position group, currently head+left_wrist+right_wrist.
• device_id: selected co-located IMU unit, imu_1202_1 or imu_1202_2.
• start_time_sec, end_time_sec: Nymeria time-code boundaries in seconds for the activity window.
• start_t_ns_global, end_t_ns_global: Nymeria global time-code boundaries in nanoseconds for the activity window.
• label: activity label for the selected label taxonomy.
• label_id: zero-based label index for the selected label taxonomy.
• imu: nested list with shape [T, 18].

Channel order:

head_acc_x, head_acc_y, head_acc_z, head_gyro_x, head_gyro_y, head_gyro_z, left_wrist_acc_x, left_wrist_acc_y, left_wrist_acc_z, left_wrist_gyro_x, left_wrist_gyro_y, left_wrist_gyro_z, right_wrist_acc_x, right_wrist_acc_y, right_wrist_acc_z, right_wrist_gyro_x, right_wrist_gyro_y, right_wrist_gyro_z

The exported IMU values are the processed 60 Hz benchmark windows before model-specific normalization. Shorter-than-5-second windows are not zero-padded in the Parquet files. Windows longer than 300 timesteps are split into non-overlapping 300-timestep segments, matching the AnyMo Bench evaluation construction.

Mapping Back to Nymeria

Each row includes the Nymeria session_id and global time-code fields (start_t_ns_global, end_t_ns_global). These fields let users map AnyMo Bench labels back to the original Nymeria sequence and align the labels with other Nymeria modalities, including RGB video and body motion. This means the curated fine-grained activity labels can also be used to build video-based or body-motion-based recognition tasks on top of the original Nymeria data. The official Nymeria tools are available at facebookresearch/nymeria_dataset.

A typical alignment workflow is:

from pathlib import Path

from datasets import load_dataset

NYMERIA_ROOT = Path("/path/to/nymeria")

benchmark = load_dataset(
    "CRUISEResearchGroup/AnyMo-Bench",
    "AnyMo-Bench-150-US",
)

row = benchmark["train"][0]

sequence_root = NYMERIA_ROOT / row["session_id"]
start_t_ns = int(row["start_t_ns_global"])
end_t_ns = int(row["end_t_ns_global"])
label = row["label"]

# The label applies to the full Nymeria TIME_CODE interval below.
# Use the official Nymeria tools to load video, body motion, or other modalities
# from `sequence_root`, then keep modality samples whose TIME_CODE timestamps
# fall inside [start_t_ns, end_t_ns].
activity_interval = (start_t_ns, end_t_ns)

This interval-level mapping preserves the AnyMo Bench window setting: the activity label is assigned to the whole window, so downstream video or body-motion clips should be selected over the same start and end time-code range.

Split Summary

Config	Train rows	Test rows	Train device	Test device
AnyMo-Bench-150-US	123,874	30,965	imu_1202_1	imu_1202_1
AnyMo-Bench-150-USCD	123,874	30,965	imu_1202_1	imu_1202_2
AnyMo-Bench-50-US	123,874	30,965	imu_1202_1	imu_1202_1
AnyMo-Bench-50-USCD	123,874	30,965	imu_1202_1	imu_1202_2

License and Source Dataset

AnyMo Bench is curated from Nymeria and follows the Nymeria non-commercial research-use terms. Please cite the Nymeria dataset and AnyMo when using this data.

@article{chen2026anymo,
  title={AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild},
  author={Chen, Baiyu and Li, Zechen and Wongso, Wilson and Li, Lihuan and Lin, Xiachong and Xue, Hao and Tag, Benjamin and Salim, Flora},
  journal={arXiv preprint arXiv:2605.22715},
  year={2026}
}

@inproceedings{ma2024nymeria,
  title={Nymeria: A massive collection of multimodal egocentric daily motion in the wild},
  author={Ma, Lingni and Ye, Yuting and Hong, Fangzhou and Guzov, Vladimir and Jiang, Yifeng and Postyeni, Rowan and Pesqueira, Luis and Gamino, Alexander and Baiyya, Vijay and Kim, Hyo Jin and others},
  booktitle={European Conference on Computer Vision},
  pages={445--465},
  year={2024},
  organization={Springer}
}