EgoWrist Gesture 30

Dual-View Wrist-Worn RGB–IMU Dataset for Dynamic Gesture Recognition

Peizhe Zhao¹ Zhenyang Gong² Ye Tao³ Rui Shi⁴ Nenglun Chen^5* Yang Zhang⁶ Peibei Cao⁵ Zhenglong Ding⁵ Feng Jiang⁵

¹University College London ²Shenzhen University ³Rice University ⁴Anhui Jianzhu University ⁵Nanjing University of Information Science and Technology ⁶Hubei University of Technology ^*Corresponding author

Paper Access Dataset GitHub Repository

All 30 gestures shown with dorsal and palm wrist-camera views — One annotated example from every gesture class, shown from the dorsal and palm wrist-camera views.

30Gesture classes

13Participants

336Recordings

2,751Annotated intervals

4,255Balanced segments

Dynamic gestures from the wrist

EgoWrist-Gesture30 captures complementary visual and inertial evidence for mobile, natural hand interaction.

Abstract

Wrist-worn cameras enable always-available gesture input, but observe the hand from a moving and partially occluded viewpoint. EgoWrist-Gesture30 provides two complementary RGB streams oriented toward the dorsal and palm sides of the hand, together with WT61C-TTL inertial measurements. The dataset contains 30 dynamic gesture classes performed by 13 participants across 336 indoor and outdoor recordings. Canonical frame-level temporal annotations and balanced derived segments support gesture recognition, cross-user evaluation, reliability-aware multi-view learning, and multimodal sensor-fusion research.

Dorsal RGB

A wrist-mounted camera observes the back of the hand and captures complementary motion and articulation cues.

Palm RGB

A second wrist-mounted camera observes the palm side, reducing self-occlusion through a complementary viewpoint.

Inertial sensing

WT61C-TTL measurements provide acceleration, angular velocity, and Euler-angle signals for multimodal research.

Thirty dynamic interaction primitives

The vocabulary spans finger articulation, fist and palm transitions, directional motion, symbolic actions, and whole-hand gestures.

Illustrations of the 30 dynamic gesture classes — Gesture vocabulary and representative motion directions.

Data distribution

The balanced representation contains 4,255 temporal segments. Original intervals remain the canonical annotation source.

Balanced segment count for each gesture class — Balanced segment count by gesture class.

Average balanced segment duration for each gesture class — Mean balanced-segment duration by gesture class.

Dual-view wrist-worn device

Two 160-degree fixed-focus cameras capture the dorsal and palm views at a nominal 12 FPS. A shared polling loop provides software-based approximate synchronization.

EgoWrist dual-camera device and software synchronization mechanism

CameraHBVCAM-1466 V22

Lens160° fixed focus · 1.66 mm

Frame rate12 FPS nominal

IMUWT61C-TTL

Research-ready annotations and tools

The GitHub repository hosts temporal annotations, acquisition software, processing utilities, quality-review tools, and integrity validators. Recorded RGB and IMU data are distributed separately.

dataset-root

subjectXX/
└── sceneY/
    └── recording_id/
        ├── dorsal/
        │   ├── 0.jpg
        │   └── ...
        ├── palm/
        │   ├── 0.jpg
        │   └── ...
        └── sensor/
            └── sensor.csv

Inclusive temporal intervals

Each recording CSV stores begin, end, and zero-based label fields. Both frame boundaries are inclusive.

14,27,1014 frames · class 10

No prescribed split

The release does not impose training, validation, or test partitions. Participant-disjoint protocols are recommended for cross-user evaluation.

Access the dataset

Annotations and research tools are versioned on GitHub. Submit the access request form for the separately distributed recorded RGB and IMU data.

Annotations

Canonical per-recording CSV files and aggregate JSON representations for all 30 classes.

Browse annotations →

Recorded data

Dorsal RGB, palm RGB, and WT61C-TTL measurements are available by request for non-commercial academic and scientific research.

Request access →

Tools

Multimodal acquisition, annotation processing, visualization, review, and validation utilities.

Browse tools →

Accompanying research

UNDER PEER REVIEW

EgoWrist: Reliability-Aware and Biomechanically Grounded Dynamic Gesture Recognition

Peizhe Zhao, Zhenyang Gong, Ye Tao, Rui Shi, Nenglun Chen, Yang Zhang, Peibei Cao, Zhenglong Ding, Feng Jiang

The paper introduces a dual-view recognition framework that combines hierarchical feature extraction, biomechanical prototype manifold projection, and Bayesian confidence filtering for reliable wrist-worn gesture recognition.

Cite EgoWrist

Please cite the accompanying submission when using EgoWrist-Gesture30 in academic work.

@unpublished{zhao2026egowrist,
  title  = {EgoWrist: Reliability-Aware and
            Biomechanically Grounded Dynamic
            Gesture Recognition},
  author = {Zhao, Peizhe and Gong, Zhenyang and
            Tao, Ye and Shi, Rui and Chen, Nenglun and
            Zhang, Yang and Cao, Peibei and
            Ding, Zhenglong and Jiang, Feng},
  year   = {2026},
  note   = {Manuscript under peer review}
}