OpenWatch: A Multimodal Benchmark for Hand Gesture Recognition on Smartwatches

OpenWatch is a multimodal wrist-worn sensor dataset for hand gesture recognition. It captures 59 discrete hand gestures using a custom smartwatch equipped with:

• Photoplethysmography (PPG)
• 3-axis accelerometer
• 3-axis gyroscope

Gesture Demonstrations

🎬 Full gesture playlist on YouTube — demo videos for all 59 gesture classes.

Double Clench	Pinch Down	Pinch Up	Slide

Dataset Splits

Base Split — open_watch_without_augmentations

• 50 real participants (id_1 – id_50)
• Body positions vary per participant: sitting, standing, standing_arm_down, walking
• Each session contains one CSV file per gesture and a feedback.json with demographic and usability data

Augmented Split — open_watch_augmentations

• 85 participants total: id_1–id_50 are the same real participants as the base split; id_51–id_85 are motion-augmented participants
• Augmented participants extend position coverage to include walking, walking_sitting, standing_arm_down, and walking_standing_arm_down
• Walking augmentation is derived from real walking sequences (walking_sequences_for_augmentation.csv)

Gesture Classes

The dataset contains 59 gesture classes:

Category	Gestures
Basic	clench, extend, flex, pinch, pinky_pinch, slide, snap, spread
Directional	deviate_in, deviate_out, pinch_down, pinch_left, pinch_right, pinch_up, rotate_in, rotate_out
Double	double_clench, double_pinch, double_pinky_pinch, double_snap, double_spread
Hand shape	index_pointing, peace, spiderman, thumbs_down, thumbs_up, vulcan_salute
Activities of Daily Living	air, answer_phone, balling, bill_please, cat_grab, cheering_fist, clapping, finger_up, good_luck, grab_cup, handshake, horns, italian_pinch, love_fingers, love_hands, money_sign, no_waving, nock_on_wood, ok_sign, open_palm, peace_up, pistol_gun, power_sign, small_amount, steepling, stop_sign, type_computer, type_phone, waiting_bored, wash_hands, waving_hello, write_pen

File Structure

open_watch_without_augmentations/
  label_map.json                        # gesture name → class ID (0–58)
  analysis/
    watch_delta_shifts.json             # per-participant temporal alignment offsets
  id_<N>/
    <position>/
      <gesture>.csv                     # sensor time-series (PPG_mean, AccX/Y/Z, GyroX/Y/Z)
      feedback.json                     # demographics + per-gesture usability ratings

open_watch_augmentations/
  label_map.json
  walking_sequences_for_augmentation.csv
  id_<N>/
    <position>/
      <gesture>.csv
      feedback.json                     # only for real participants (id_1–id_50)

videos/
  demo_double_clench.mp4
  demo_pinch_down.mp4
  demo_pinch_up.mp4
  demo_slide.mp4

Sensor CSV Format

Each gesture CSV contains one row per sensor sample:

Column	Description
PPG_mean	Photoplethysmography (mean across channels, raw sensor counts)
AccX, AccY, AccZ	3-axis accelerometer (raw integer units)
GyroX, GyroY, GyroZ	3-axis gyroscope (raw integer units)

Feedback JSON Format

{
  "id": "1",
  "age": "26",
  "gender": "male",
  "handedness": "left",
  "position": "sitting",
  "experience": "often",
  "globalRating": 10,
  "globalFeedback": "...",
  "gestureFeedback": {
    "Clench": { "easiness": 10, "usability": 10 },
    ...
  },
  "gestureTimestamps": [...]
}

Loading the Data

import pandas as pd
import json
from pathlib import Path

root = Path("open_watch_without_augmentations")

# Load a single gesture recording
df = pd.read_csv(root / "id_1" / "sitting" / "clench.csv")

# Load label map
with open(root / "label_map.json") as f:
    label_map = json.load(f)

# Load participant feedback
with open(root / "id_1" / "sitting" / "feedback.json") as f:
    feedback = json.load(f)

Citation

@article{bonazzi2025openwatch,
  title   = {OpenWatch: A Multimodal Benchmark for Hand Gesture Recognition on Smartwatches},
  author  = {Bonazzi, Pietro and Ahmed, Youssef and Eckert, Daniel and Ronco, Andrea and Zeng, Junjie and Magno, Michele and Dai, Dengxin},
  journal = {NeurIPS (under review)},
  year    = {2025}
}

License

This dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0).