The ClassyGlass Dataset contains multimodal time-series data collected from wearable smart glasses used by real participants. The dataset includes measurements from multiple onboard sensors, including an accelerometer, gyroscope, magnetometer, and pressure sensor. The data are organized by subject and recording session to support reproducible research and machine learning workflows.
ClassyGlass is a comprehensive wearable sensing dataset designed for activity recognition and behavioral analysis research. It provides temporally aligned sensor streams recorded during controlled and semi-naturalistic sessions.
Figure 1: The ClassyGlass hardware setup with MetaMotionC sensor attached to the right temple.
Data were collected using the MetaMotionC wearable Inertial Measurement Unit (IMU) sensor Sensor Link. The device provides real-time and continuous motion and environmental sensing through an integrated 9-axis IMU and a pressure sensor. Data are transmitted via Bluetooth Low Energy (BLE) using an open-source API. Onboard Kalman filter–based sensor fusion is applied to improve signal quality and reduce noise. All sensor data are timestamped to enable precise synchronization across data streams.
Figure 2: The ClassyGlass hardware setup, showing the internal view of the MetaMotionC 10-axis IMU.
| Sensor | Measurement Range | Resolution | Sampling Rate |
|---|---|---|---|
| Accelerometer | ±2, ±4, ±8, ±16 g | 16-bit | 0.001–100 Hz (stream), up to 800 Hz (log) |
| Gyroscope | ±125, ±250, ±500, ±1000, ±2000 °/s | 16-bit | 0.001–100 Hz (stream), up to 800 Hz (log) |
| Magnetometer | ±1300 µT (x,y), ±2500 µT (z) | 0.3 µT | 0.001–25 Hz |
| Barometer/Pressure/Altimeter | 300–1100 hPa | 0.01 hPa | 0.001–50 Hz |
Each dataset contains its own data description sheet. Follow that for dataset specific information
ClassyGlass/
- Datasets/
- Dataset_1A (12 users with 11 activities each)
- Dataset_1B (15 users with 15 activities each)
- Dataset_2 (35 users with 6 activities each)
- Data description 2
- TUG_Annotations (These videos can be used to annotate the TUG task. We provide example annotations for User 22 and User 23)
File name: Each data file follows a structured naming convention that encodes metadata about
the device, recording time, sensor type, and sampling configuration.
<experiment_id>_<device>_<timestamp>_<device_id>_<sensor_type>_<sampling rate>Hz_<firmware_version>.csv
Columns include:
timestamp— Time of measurement (ISO 8601 or Unix epoch in milliseconds)elapsed- Time relative to starting timex-axis,y-axis,z-axis— 6-axis Accelerometer sensor, Linear acceleration (m/s²)x-axis,y-axis,z-axis— 6-axis Gyroscope sensor, Angular velocity (deg/s)x-axis,y-axis,z-axis— 3-axis Magnetometer sensor, Magnetic field strength (T)pressure— Barometer/Pressure/Altimeter, Atmospheric pressure (hPa)
The following table provides a comparative overview of the frame counts and total durations for both data subsets.
| Sensor Type | Dataset_1A | Dataset_1B | Dataset_2 (Discrete) | Dataset_2 (Continuous TUG) | Total |
|---|---|---|---|---|---|
| Accelerometer (100Hz) | 1,504,164 | 2,982,352 | 2,057,652 | 315,682 | 6,859,850 |
| Gyroscope (100Hz) | 1,504,592 | 2,971,744 | 2,058,286 | 315,820 | 6,850,442 |
| Magnetometer (20Hz) | 296,744 | 592,144 | 410,672 | 62,994 | 1,362,554 |
| Pressure (7.33Hz) | 109,450 | 219,583 | 151,119 | 22,892 | 503,044 |
| Duration | 4h 10m 40s | 8h 18m 19s | 5h 42m 54s | 0h 52m 40s | 19h 04m 33s |
Following table contains demographic informations
| Characteristic | Dataset 1 | Dataset 2 | Total |
|---|---|---|---|
| Count (N) | 27 | 35 | 62 |
| Gender | 18M / 9F | 25M / 10F | 43M / 19F |
| Age Range | 21–61 | 21–65 | 21–65 |
| Glasses Wearers | 60% | ~62% | ~61% |
| Category | Activity Description |
|---|---|
| Sedentary | Sitting & Reading (Book) |
| Sitting & Writing (Notebook) | |
| Computer: Typing | |
| Computer: Browsing | |
| Natural Fidgeting (Head/Body) | |
| Transitional | Moving Chair / Adjusting |
| Sit-to-Stand Transition | |
| Pick up items from floor (Sitting) | |
| Pick up items from floor (Standing) | |
| Ambulatory | Standing Still |
| Walking | |
| Running | |
| Taking Stairs | |
| Donning/Doffing | Wear/Remove Device (Sitting) |
| Wear/Remove Device (Standing) | |
| Timed up and go test (TUG) | Sit - Stand - Up - Walk to a point - turn - walk back to chair - sitdown |
Figure 3: Schematic of the Timed Up and Go (TUG) test. The TUG path and the four activities performed during a TUG. These transitions are: 1) sitting, 2) sit-to-stand 2) walking- out - turning - walking-in, 4) stand-to-sit. The continuous sequence challenges models to detect transitional boundaries e.g., the exact moment of ’Turn’) within a continuous stream.
(a) Sitting and Reading | 
(b) Working on Computer | 
(c) Sit-to-Stand Transition |
(d) Walking | 
(e) Picking Up Items | 
(f) Video Labeling (Ground Truth) |
Figure 4: Data visualization of the ClassyGlass benchmark. (a-e) Multi-modal sensor fingerprints for distinct activities, showing 4 sensor streams (Accel, Gyro, Pressure, Mag). (f) An example of the high-fidelity synchronization between the video ground truth and accelerometer peaks used for validating the dataset labels.
Category matching using mod of number ot the experiments
from pathlib import Path
import pandas as pd
file_path = "Datasets/Dataset_1A/User1/1_MetaWear_2019-09-14T13.41.43.335_F1E55E2FE95F_Accelerometer_100.000Hz_1.4.5.csv"
# Load CSV
df = pd.read_csv(file_path)
# Extract filename
filename = Path(file_path).name
# Parse first number before underscore
raw_exp_num = int(filename.split("_")[0])
# Mod with 11 (map 1 -> 11)
experiment_number = raw_exp_num % 11 or 11
if experiment_number == 1:
experiment = "Reading a book"
print("Raw experiment number:", experiment_number)
print("Mapped experiment:", experiment)- Visualizations and feature extraction
- Classifications
- 🧑💻 The code in this repository is licensed under the MIT License.
- 📊 The datasets in the
data/directory are released under the CC BY 4.0 License.