istLab/edgelet-imu-gestures

edgelet IMU Gestures

9-axis IMU recordings of six hand gestures, made with an Arduino Nano 33 BLE. This is the training data for the MLIoT course labs (edgelet); it is the same data as edgelet's built-in base dataset.

Classes (6): idle, circle, leftright, updown, snake, twist

Channels (9): accX accY accZ (m/s²), gyrX gyrY gyrZ (deg/s), magX magY magZ (µT)

Sampling rate: 100 Hz. Each recording is about 10 s.

How the data was split

1. Test set withheld. The full collection has 149 recordings. 30 of them (5 per class) are the held-out test set of the course leaderboard and are not published.
2. Stratified by class and variant. Each class was recorded in several variants (the way the gesture is performed, e.g. a large circle or a small wrist circle; each variant usually in five board orientations). The test and validation recordings are spread over the variants, so every variant appears in training, but no recording appears in more than one split.
3. Train / validation split by recording. Of the remaining 119 recordings, 4 per class go to validation and the rest to train (fixed seed). All windows from one recording stay in the same split, so validation accuracy is not inflated by near-identical windows.
4. 3 s windows (windows config): non-overlapping windows of 300 samples starting at sample 0, 300 and 600.

class	train	validation	test (withheld)
idle	16	4	5
circle	16	4	5
leftright	16	4	5
updown	16	4	5
snake	16	4	5
twist	15	4	5
total	95	24	30

splits.csv lists every published recording with its class, variant and split.

Configurations

• windows (default): one row per 3 s window. Columns: recording_id, label, window_index, start_sample, and the 9 channel columns, each a list of 300 floats.
• raw: one row per full recording, for your own windowing (e.g. overlapping windows for training). Columns: recording_id, label, variant, split, device, n_samples, and the 9 channel columns. If you make your own windows, cut them after splitting.

Usage

from datasets import load_dataset
import numpy as np

ds = load_dataset("istLab/edgelet-imu-gestures")          # windows
AX = ["accX","accY","accZ","gyrX","gyrY","gyrZ","magX","magY","magZ"]
X = np.stack([np.stack([r[a] for a in AX], -1) for r in ds["train"]])   # (windows, 300, 9)
y = ds["train"]["label"]

raw = load_dataset("istLab/edgelet-imu-gestures", "raw")   # full recordings