Oceanveo/meta-quest-ego

Meta Quest Ego Dataset

This dataset contains 14 short egocentric capture episodes recorded from a Meta Quest-style headset. Each episode includes synchronized stereo camera video, headset pose, camera calibration, per-camera pose tracking, and a Vuer-compatible tracking stream with hand landmarks.

This README describes the dataset structure and sensor streams included in each episode.

Dataset Layout

.
+-- episode_001/
|   +-- videos/
|   |   +-- left_camera.mp4
|   |   +-- right_camera.mp4
|   +-- camera/
|   |   +-- left_camera_image_format.json
|   |   +-- right_camera_image_format.json
|   |   +-- left_camera_characteristics.json
|   |   +-- right_camera_characteristics.json
|   +-- tracking/
|       +-- hmd_poses.csv
|       +-- camera_pose_tracking.jsonl
|       +-- tracking_vuer_compat.jsonl
+-- episode_002/
+-- ...

Each episode_XXX directory follows the same file naming scheme.

High-Level Contents

• Episodes: 14
• Total size: about 20 GB
• Video files: 28 MP4 files, one left and one right camera stream per episode
• Tabular pose files: 14 CSV files
• Tracking streams: 28 JSONL files
• Camera metadata files: 56 JSON files

File Meanings

videos/left_camera.mp4 and videos/right_camera.mp4

Stereo egocentric camera streams for the left and right headset cameras. The files are MP4 containers with H.264 video. Camera resolution is 1280 x 1280 across the dataset.

tracking/hmd_poses.csv

Head-mounted display pose samples. Columns:

• unix_time: Unix epoch timestamp in milliseconds.
• ovr_timestamp: Oculus/OVR runtime timestamp in seconds.
• pos_x, pos_y, pos_z: headset position in the tracking coordinate frame.
• rot_x, rot_y, rot_z, rot_w: headset orientation quaternion in x, y, z, w order.

tracking/camera_pose_tracking.jsonl

Per-camera pose tracking stream. Each line is a JSON object. Key fields:

• t_unix_ms: Unix epoch timestamp in milliseconds.
• side: left or right, identifying the camera stream.
• ovr_time_sec: Oculus/OVR runtime timestamp in seconds.
• ovr_time_source: timestamp source, e.g. camera_monotonic.
• head.matrix_col_major_4x4: headset pose as a 4 x 4 homogeneous transform flattened in column-major order.
• characteristics.translation: camera translation relative to the headset/gyro reference frame.
• characteristics.rotation_xyzw: camera orientation quaternion in x, y, z, w order.
• camera.matrix_col_major_4x4: camera pose as a 4 x 4 homogeneous transform flattened in column-major order.

The side field should be used when aligning these rows with the left or right video.

tracking/tracking_vuer_compat.jsonl

Merged tracking stream formatted for Vuer-style visualization or replay. Each line is a JSON object. Key fields:

• t_unix_ms: Unix epoch timestamp in milliseconds.
• ovr_time_sec: Oculus/OVR runtime timestamp in seconds.
• left_hand_sample_time_sec, right_hand_sample_time_sec: hand sample timestamps.
• camera.matrix_col_major_4x4: camera/head pose transform flattened in column-major order.
• leftHand, rightHand: 4 x 4 hand transforms flattened in column-major order.
• leftLandmarks, rightLandmarks: 25 3D hand landmark points per hand.

All-zero landmark rows represent unavailable hand tracking for that sample.

camera/left_camera_image_format.json and camera/right_camera_image_format.json

Per-camera image acquisition format. Key fields:

• width, height: camera image dimensions.
• format: image format at capture time.
• baseTime: monotonic-to-Unix timestamp anchor for aligning camera frames with tracking streams.

camera/left_camera_characteristics.json and camera/right_camera_characteristics.json

Per-camera calibration and device metadata. Key fields:

• cameraId: device camera identifier.
• cameraPositionId: side/position identifier, with 0 for left and 1 for right.
• pose.translation: camera translation relative to the headset/gyro reference frame.
• pose.rotation: camera orientation quaternion in x, y, z, w order.
• pose.reference: pose reference frame, here GYROSCOPE.
• intrinsics: pinhole camera parameters fx, fy, cx, cy, and skew.
• distortion: distortion coefficient field.
• sensor: physical and pixel-array metadata.

Time and Synchronization

The dataset provides both wall-clock and runtime timestamps:

• CSV pose files use unix_time and ovr_timestamp.
• JSONL streams use t_unix_ms and ovr_time_sec.
• Camera image format files contain a monotonic-to-Unix anchor in baseTime.

For synchronization, align samples by nearest timestamp (ovr_timestamp/ovr_time_sec or Unix milliseconds) rather than by row number. Video frame counts and tracking row counts differ by stream and episode.

Coordinate and Transform Notes

• Quaternions are stored as x, y, z, w.
• Matrix fields named matrix_col_major_4x4 contain 16 values in column-major order.
• Positions and transforms are represented in the headset/OVR tracking frame.
• Translation entries in the 4 x 4 matrices appear in the final column when interpreted as column-major homogeneous transforms.