ravenwing/cheedeh-IMU-data

Cheedeh IMU Gesture Data

Phone IMU recordings of mid-air gestures drawn in the air with the phone. Collected to train a real-time gesture classifier running on Android.

Gestures (labels)

Label	Gesture
z	Letter Z
m	Letter M
s	Letter S
o	Letter O (circle)
none	No gesture / idle

Splits

Split	Samples	m	none	o	s	z
train	266	31	111	43	42	39
test	54	15	20	8	4	7

Data Format

Each sample is a separate JSON file named gesture_{label}_{YYYYMMDD}_{HHMMSS}.json.

{
  "gesture_id": "m_20260210_160623",
  "gesture_name": "m",
  "timestamp": "2026-02-10T15:06:23.660296Z",
  "duration_ms": 3246,
  "sample_rate_hz": 57,
  "data": [
    {
      "t":   1,
      "x":  -0.072,  "y": -0.143,  "z":  0.456,
      "gx": -0.011,  "gy": -0.043, "gz":  0.020,
      "grx": 0.522,  "gry": 4.386, "grz": 8.759,
      "qx":  0.152,  "qy": -0.175, "qz": -0.653, "qw": 0.721
    },
    ...
  ]
}

Fields in data array

Field	Sensor	Unit
t	Elapsed time	ms
x, y, z	Linear acceleration (TYPE_LINEAR_ACCELERATION)	m/s²
gx, gy, gz	Gyroscope (angular velocity)	rad/s
grx, gry, grz	Gravity vector (rotation frame)	m/s²
qx, qy, qz, qw	Orientation quaternion	—

Sample rate is approximately 50–57 Hz. Duration varies by gesture (typically 1–4 s).

Loading

from datasets import load_dataset

ds = load_dataset("ravenwing/cheedeh-IMU-data")
sample = ds["train"][0]
# sample["gesture_name"]  → "m"
# sample["data"]          → list of dicts with t, x, y, z, ...

Collection

Recorded with an Android app (gather-data) using TYPE_LINEAR_ACCELERATION and gyroscope sensors. Each session was a manual recording of one gesture, labeled at collection time.

Related

• Training code: learn/ — SVM/RF classifier achieving ~0.76 test accuracy, exported to ONNX for on-device inference.