Delta-Intelligence/OmniEgo

DOHC Headset Egocentric Whole-body Dataset

DOHC is a headset multi-camera human motion data collection dataset designed for humanoid intelligence, embodied AI, whole-body motion understanding, and imitation learning.

Overview

The DOHC dataset is exported from the DOHC headset multi-camera human motion capture system developed by Delta Intelligence. Each recorded episode contains synchronized multi-view video streams and whole-body skeleton / headset pose data.

The dataset is designed to support research and development in:

• Human behavior observation and understanding
• Human detection, keypoint detection, and pose estimation
• Headset device trajectory and motion analysis
• Egocentric visual observation and synchronization checking
• Whole-body motion modeling for humanoid robots
• Imitation learning, behavior cloning, and embodied intelligence model training
• Visual-inertial odometry / SLAM related observation analysis

Unlike purely third-person datasets, DOHC combines egocentric RGB / fisheye views with global whole-body skeleton and headset pose states, making it suitable for learning from human demonstrations in realistic environments.

---

Headset Device

The headset device is designed for full-body human data acquisition from an egocentric perspective.

Device Parameters

Parameter	Description
Processor	Rockchip RK3588-class processor
Memory / Storage	4GB + 32GB / 8GB + 64GB configurations
Connectivity	Dual-band Wi-Fi, Bluetooth 5.3, Type-C interface
Audio	Dual microphones and speaker-supported audio collection
Battery	7.4V 1000mAh
Charging	Dedicated charging base
Tracking Accuracy	Full-body system spatial positioning accuracy within approximately 2 cm
Data Types	RGB image, IMU data, whole-body skeleton data, microphone audio, spatial pose trajectory

---

Dataset Structure

The dataset is distributed as MCAP files converted from two source formats: processed H5 files and JPEG session folders. Each MCAP stores synchronized image streams and structured JSON messages under /dohc/* topics.

Contents

Split	Files	Frames
h5	6 MCAP files	1,238 frames
jpeg	13 MCAP files	5,248 frames

Files

.
├── h5/
│   └── *.mcap           # One MCAP per processed H5 file
├── jpeg/
│   └── *.mcap           # One MCAP per JPEG session folder
└── manifest.jsonl       # Source path, output path, frame count, and image counts per MCAP

• h5/*.mcap: one MCAP file converted from each processed H5 source file.
• jpeg/*.mcap: one MCAP file converted from each JPEG session folder.
• manifest.jsonl: source path, output path, frame count, and image counts per MCAP.

MCAP Topic Layout

Each MCAP contains a subset of the following /dohc/* topics:

Topic	Encoding	Description
/dohc/metadata	JSON	Episode-level metadata. Usually one message per MCAP.
/dohc/state	JSON	Frame-aligned headset state, including pose and motion information.
/dohc/skeleton	JSON	Frame-aligned whole-body skeleton data. Present in JPEG-session MCAPs in the provided sample.
/dohc/cam0/jpeg	JPEG	Main front-facing RGB camera stream.
/dohc/cam1/jpeg	JPEG	Left-front RGB / wide-angle camera stream.
/dohc/cam2/jpeg	JPEG	Right-front RGB / wide-angle camera stream.
/dohc/t265_left/jpeg	JPEG	Left grayscale tracking / fisheye camera stream.
/dohc/t265_right/jpeg	JPEG	Right grayscale tracking / fisheye camera stream.

Episode Modalities

Conceptually, each episode provides:

1. Visual observations

   • Main RGB view: /dohc/cam0/jpeg
   • Side RGB / wide-angle views: /dohc/cam1/jpeg, /dohc/cam2/jpeg
   • Tracking / fisheye views: /dohc/t265_left/jpeg, /dohc/t265_right/jpeg

2. Structured state and annotation data

   • Episode metadata: /dohc/metadata
   • Headset pose and motion state: /dohc/state
   • Whole-body skeleton: /dohc/skeleton, when available

---

Video Streams

All image streams are stored as JPEG messages in MCAP topics. The topic names below correspond to the Episode Modalities section.

/dohc/cam0/jpeg — Main RGB Camera

cam0 is the main front-facing RGB camera stream. It is usually used as the primary visual input for human observation, human detection, keypoint detection, pose estimation, and action segmentation.

Field	Value
Topic	/dohc/cam0/jpeg
Type	RGB
Frame Rate	30 FPS
Resolution	1920 × 1080

Typical usage:

• Human action visualization
• Human detection and keypoint detection
• Pose estimation model input
• Synchronization check with 3D skeleton data

Camera parameters:

Camera matrix:
[[283.936956,   0.      , 953.207147],
 [  0.      , 284.022358, 532.647632],
 [  0.      ,   0.      ,   1.      ]]

Distortion coefficients:
[-0.02257256, 0.00926648, -0.0111663, 0.00267249]

---

/dohc/cam1/jpeg / /dohc/cam2/jpeg — Left and Right RGB / Wide-angle Cameras

cam1 and cam2 are left-front and right-front RGB / wide-angle camera streams. They provide additional viewpoints for multi-view human observation and scene understanding.

Field	Value
Topics	/dohc/cam1/jpeg, /dohc/cam2/jpeg
Type	RGB
Frame Rate	30 FPS
Resolution	1280 × 720

Typical usage:

• Complementing the main RGB camera
• Multi-view human detection
• Visibility checking under different viewpoints
• Action understanding and post-processing
• Egocentric multi-view observation

---

/dohc/t265_left/jpeg / /dohc/t265_right/jpeg — Left and Right Tracking Cameras

t265_left and t265_right are grayscale fisheye-style tracking camera streams. They are mainly used for visual odometry, SLAM-related visual feature analysis, headset motion observation, and synchronization checking.

Field	Value
Topics	/dohc/t265_left/jpeg, /dohc/t265_right/jpeg
Type	Grayscale
Frame Rate	30 FPS
Resolution	848 × 800

Typical usage:

• VIO / SLAM visual feature analysis
• Headset device motion observation
• Synchronization check with headset pose data
• Fisheye vision post-processing and debugging

Camera parameters:

Camera matrix:
[[426.11546713,   0.        , 657.06272437],
 [  0.        , 427.85036814, 355.11072595],
 [  0.        ,   0.        ,   1.        ]]

Distortion coefficients:
[0.18449862, -0.13104704, 0.0027557, -0.00128653, -0.03698853]

---

Skeleton and State Data

DOHC stores whole-body skeleton and headset motion information as MCAP topics instead of standalone NPZ files.

Topic	Description
/dohc/metadata	Per-episode metadata, including source information and recording-level attributes
/dohc/state	Headset state, including translation, rotation, velocity, angular velocity, and acceleration
/dohc/skeleton	Frame-aligned whole-body skeleton data

/dohc/skeleton — Whole-body Skeleton

The skeleton topic provides frame-aligned 3D whole-body joint positions using a 24-joint SMPL-style layout.

Field	Value
Topic	/dohc/skeleton
Skeleton Layout	SMPL-style 24-joint body skeleton
Coordinate Format	3D joint coordinates [x, y, z]
Shape Concept	(frame_num, 24, 3)
Alignment	Frame-aligned with image topics

The joint order is:

Index	Joint Name
0	pelvis
1	left_hip
2	right_hip
3	spine1
4	left_knee
5	right_knee
6	spine2
7	left_ankle
8	right_ankle
9	spine3
10	left_foot
11	right_foot
12	neck
13	left_collar
14	right_collar
15	head
16	left_shoulder
17	right_shoulder
18	left_elbow
19	right_elbow
20	left_wrist
21	right_wrist
22	left_hand
23	right_hand

Common skeleton connections:

SMPL_BONES = [
    (0, 1), (0, 2), (0, 3),
    (1, 4), (4, 7), (7, 10),
    (2, 5), (5, 8), (8, 11),
    (3, 6), (6, 9), (9, 12), (12, 15),
    (12, 13), (13, 16), (16, 18), (18, 20), (20, 22),
    (12, 14), (14, 17), (17, 19), (19, 21), (21, 23),
]

Typical usage:

• Whole-body motion visualization
• Video-skeleton synchronization checking
• Motion modeling and motion prediction
• Imitation learning and humanoid motion retargeting

/dohc/state — Headset State

The state topic describes the headset pose and motion state in the global coordinate system.

Field	Shape Concept	Description
pose_translation	(frame_num, 3)	Headset global translation [x, y, z]
pose_rotation	(frame_num, 4)	Headset global rotation quaternion [qx, qy, qz, qw]
pose_velocity	(frame_num, 3)	Headset linear velocity [vx, vy, vz]
pose_angular_velocity	(frame_num, 3)	Headset angular velocity [wx, wy, wz]
pose_acceleration	(frame_num, 3)	Headset linear acceleration [ax, ay, az]

These fields can be used for headset trajectory analysis, motion smoothness checking, sensor-fusion research, and alignment with the image and skeleton topics.

Frame Alignment

All topics are intended to be aligned by frame index or timestamp:

/dohc/cam0/jpeg[f]
  <-> /dohc/cam1/jpeg[f]
  <-> /dohc/cam2/jpeg[f]
  <-> /dohc/t265_left/jpeg[f]
  <-> /dohc/t265_right/jpeg[f]
  <-> /dohc/skeleton[f]
  <-> /dohc/state[f]

---

Contact

For questions, feedback, or collaboration requests, please contact:

Email: sales@deltai.com

Website: www.deltai.com

WeChat Official Account:

---

Acknowledgement

This dataset is collected using the Delta Intelligence data acquisition system, designed for humanoid intelligence, embodied AI, and whole-body human motion data collection.