Research Benchmark
A Multimodal Egocentric Video and Inertial Dataset for Motor Impairment Analysis
1 University of Florence, Italy · fatemah.alhamdoosh@unifi.it, pietro.pala@unifi.it
2 Meta Reality Labs, USA · abduallah.adel.omar@gmail.com
3 University of Edinburgh, UK · D.Arvind@ed.ac.uk
Motor impairments, including tremor, bradykinesia, gait abnormalities, and postural instability, are common across many neurological and movement-related conditions. Conventional clinical assessments are often intermittent and may fail to capture subtle temporal variations in motor behavior. While wearable IMUs and third-person video have shown promise for objective motor assessment, third-person recordings raise privacy concerns and require constrained acquisition setups. In contrast, egocentric vision provides a more naturalistic and privacy-aware alternative.
In this work, we introduce EgoInertia-MI, a multimodal benchmark dataset combining synchronized egocentric video and wearable IMU signals for motor impairment analysis. The dataset contains 19 upper- and lower-body activities performed by healthy volunteers simulating varying levels of motor impairment severity levels: no impairment, mild impairment, and severe impairment. We establish two benchmark tasks: action recognition and motor impairment severity estimation, and evaluate multiple unimodal and multimodal baselines. Experimental results show that egocentric video provides strong cues for motor impairment assessment, while multimodal fusion achieves the best overall performance, reaching 0.78 Macro-F1 for severity estimation and 0.93 Macro-F1 for action recognition. These findings highlight the potential of combining egocentric vision and wearable sensing for ecologically valid and privacy-aware motor assessment.
The dataset is divided into two modalities: **Video** and **IMU**. Data are organized hierarchically by participant, session, and recording. Each participant typically completed 2–4 recording sessions, with each recording capturing a single activity performed at a specific impairment severity level. Corresponding IMU and video recordings are synchronized and share the same organizational structure.
EgoInertia-MI-video/
│
└── P0001/
└── video/
└── P0001_S1/
├── P0001_S1_V01.MP4
├── P0001_S1_V02.MP4
└── ...
└── P0001_S2/
├── P0001_S1_V01.MP4
├── P0001_S1_V02.MP4
└── ...
└── ...
└── P0002/
└── ...
└── P0003/
└── ...
EgoInertia-MI-IMU/
│
└── P0001/
└── aligned_IMU/
└── P0001_S1/
├── P0001_S1_V01_aligned.csv
├── P0001_S1_V02_aligned.csv
└── ...
└── P0001_S2/
├── P0001_S2_V01_aligned.csv
├── P0001_S2_V02_aligned.csv
└── ...
└── ...
└── P0002/
└── ...
└── P0003/
└── ...
Video: EgoInertia-MI-video/P0001/video/P0001_S1/P0001_S1_V01.MP4
Aligned IMU: EgoInertia-MI-IMU/P0001/aligned_IMU/P0001_S1/P0001_S1_V01_aligned.csv
Identifier format: PXXXX_SY_VZZ
Standardized train/validation/test splits, metadata files, and benchmark protocols are provided.
Each aligned IMU CSV contains synchronized wearable inertial signals with 19 columns:
rel_time_s,
w_acc_x, w_acc_y, w_acc_z,
w_gyro_x, w_gyro_y, w_gyro_z,
l_acc_x, l_acc_y, l_acc_z,
l_gyro_x, l_gyro_y, l_gyro_z,
c_acc_x, c_acc_y, c_acc_z,
c_gyro_x, c_gyro_y, c_gyro_zw = wrist, l = leg, c = chest.
acc = tri-axial acceleration, gyro = tri-axial angular velocity.
rel_time_s is the relative timestamp in seconds.
The annotation CSV is semicolon-separated and captures clip metadata, labels, and notes.
VidID;Path;HandInVideo;SensorsSideWristLeg;Action;Severity;ClassID;duration_s;notes
Task 1
The primary task is movement impairment severity estimation from IMU, video, or multimodal data. Severity is formulated as a three-class classification problem (levels 0–2), requiring models to understand both motion patterns and their activity context.
Task 2
The benchmark also includes action recognition across 19 daily living and clinically relevant activities. Actions are performed under varying impairment levels, requiring models to learn robust representations that generalize across diverse movement patterns and execution styles.
Evaluation Protocol
We adopt a 5-fold subject-disjoint cross-validation protocol with 12 training, 2 validation, and 3 test subjects per fold. Performance is reported using the Macro F1-score.
The EgoInertia-MI dataset is available through Zenodo under restricted access. To obtain the dataset, please submit an access request through the Zenodo record, including your name, affiliation, and a brief description of the intended research use.
Access requests are reviewed by the dataset authors. If you experience any problems with the access request process or need additional information, please contact Fatemah Alhamdoosh.
Please cite EgoInertia-MI in any publication that uses the dataset or builds on the proposed controlled data-collection protocol for studying impairment-aware movements.
@inproceedings{egoinertia_mi_2026,
title = {EgoInertia-MI: A Multimodal Egocentric Video and Inertial Dataset for Motor Impairment Analysis},
author = {Alhamdoosh, Fatemah and Pala, Pietro and Mohamed, Abduallah and Arvind, D K.},
booktitle = {ToDO},
year = {2026},
}The benchmark includes volunteer participants performing simulated impairment behaviors under approved study protocols. The design follows privacy-aware egocentric acquisition principles and is intended for research benchmarking, not direct clinical diagnosis.
EgoInertia-MI is a non-clinical benchmark and should be interpreted within that limitation. Ethical use, attribution, and adherence to the dataset license are required.
We thank all participants, annotators, and collaborators who contributed to the data collection and benchmark development process.