fatemah-alh/EgoInertia-MI

Overview

EgoInertia-MI is a multimodal dataset designed for motor impairment analysis using synchronized egocentric video and inertial sensing. The dataset captures daily activities and clinically relevant motor tasks performed under different simulated impairment conditions.

The dataset supports research on multimodal learning, action recognition, and motor impairment severity estimation from egocentric vision and wearable IMU signals.

Daily Activities and Clinical Tasks

The dataset includes 19 activities spanning motor assessment tasks and activities of daily living.

These activities cover:

• upper-limb movements;

• gait and walking-related tasks; stair ascent and descent;

• daily activities such as eating and drinking.

Each activity was performed under three simulated impairment conditions:

• Natural / no impairment

• Mild impairment

• Severe impairment

Multimodal Motion Capture

EgoInertia-MI combines egocentric video with synchronized inertial measurements.

The syncronised dataset includes:

• Chest-mounted GoPro HERO12 video recorded at 30 FPS;

• wrist and leg  IMU signals sampled at 25.5 Hz;

• chest IMU signals extracted from the camera;

Dataset Structure

The dataset is divided into two main modalities:

• Video

• IMU

Data are organized hierarchically by participant, session, and recording. Each participant typically completed 2–4 recording sessions, with each recording corresponding to a single activity performed at a specific impairment severity level.

The video and IMU recordings are synchronized and follow a consistent naming structure.

Example Paths

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

Recordings follow the identifier format:

PXXXX_SY_VZZ

where:

• PXXXX = participant ID;

• SY = session ID;

• VZZ = video/recording ID.

Standardized train/validation/test splits, metadata files, and benchmark protocols are also provided.

IMU Data Format

Each aligned IMU CSV file contains synchronized 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_z

The prefixes indicate the sensor source:

• w_ = wrist IMU;

• l_ = leg IMU;

• c_ = chest/camera IMU.

The signals include tri-axial accelerometer and gyroscope measurements for each sensor.

Annotations

The annotation file is semicolon-separated and contains metadata, labels, and optional notes for each recording.

Annotation Columns

VidID;Path;HandInVideo;SensorsSideWristLeg;Action;Severity;ClassID;duration_s;notes

Column descriptions:

• VidID: unique recording identifier, for example P0001_S1_V01;

• Path: relative video file path;

• HandInVideo: visible hand in the egocentric video, for example R, L, or RL;

• SensorsSideWristLeg: side of the wearable wrist and leg sensors, for example RR or LL;

• Action: activity label, for example FingerTapping;

• Severity: simulated impairment level;

• ClassID: numeric activity class index;

• duration_s: recording duration in seconds;

• notes: optional comments.

Benchmark Tasks

Task 1: Motor Impairment Severity Estimation

The primary benchmark task is motor impairment severity estimation from IMU, video, or multimodal data.

Severity estimation is formulated as a three-class classification problem:

• level 0: no impairment;

• level 1: mild impairment;

• level 2: severe impairment.

This task requires models to capture motion patterns while also considering the activity context in which the movement is performed.

Task 2: Action Recognition

The second benchmark task is action recognition across 19 daily living and clinically relevant activities.

Since each activity is performed under varying impairment levels, models are expected to learn robust representations that generalize across different movement patterns, speeds, and execution styles.

Evaluation Protocol

We adopt a 5-fold subject-disjoint cross-validation protocol.

For each fold:

• 12 subjects are used for training;

• 2 subjects are used for validation;

• 3 subjects are used for testing.

Performance is reported using the Macro F1-score, which accounts for class imbalance and provides a balanced evaluation across classes.

Dataset Page

Additional information about the EgoInertia-MI dataset, including dataset details, benchmark information, code, and updates, is available on the project page:

EgoInertia-MI Project Page:
https://fatemah-alh.github.io/EgoInertia-MI-Page/

Please refer to the project page for the most up-to-date information about the dataset release, access instructions, code, and benchmark protocols.

 

Dataset Access

Access to the EgoInertia-MI dataset is restricted and may be granted only upon request for legitimate academic, educational, and non-commercial research purposes. Requests must be submitted using a valid institutional academic email address associated with a recognised university, research institute, or academic organisation. Requests submitted from personal or non-institutional email addresses may not be approved.

To request access to the EgoInertia-MI dataset, applicants must:

1. Submit the access request through an authenticated Zenodo user account.

Zenodo Link

2. Complete the following form. Access requests will be reviewed based on the information provided in the form.

By requesting access to the dataset, users agree to comply with the following conditions:

1. The dataset shall be used solely for academic research and educational purposes.

2. The dataset, or any part thereof, shall not be redistributed, shared, published, uploaded, or made available through any other platform or repository.

3. Users shall not attempt to identify, contact, or re-identify any participant represented in the dataset.

4. The dataset shall not be used for commercial purposes, surveillance, biometric identification, biometric recognition, or any harmful, unethical, or unlawful applications.

5. Users shall store the dataset securely and restrict access to authorised members of the approved research project only.

6. Any publication, report, or work resulting from the use of the dataset shall cite both the EgoInertia-MI paper and the corresponding Zenodo dataset record.

7. Users shall delete the dataset upon request by the dataset owners or when it is no longer required for the approved research purpose.