EgoSight-Sample-500H: Egocentric Bimanual Activity Dataset

EgoSight-Sample-500H is an open-source dataset of everyday human activity recorded from a head-mounted perspective. The release is designed to provide 500 hours of egocentric, bimanual task data across settings such as bookstores and libraries, factories, hotels, offices, pharmacies, schools, and warehouses.

Each episode brings together synchronized multi-view video, head and hand motion, inertial measurements, camera calibration, and English task descriptions. The combination makes it possible to study what a person is doing, how both hands move, and how an activity unfolds over time.

Visualize an Episode

Download an episode's .mcap file and open it in Rerun Viewer or Foxglove. These viewers provide a direct way to explore the recording's synchronized video and sensor topics in the MCAP container.

Dataset at a Glance

Dimension	Description
Open-source release duration	500 hours
Data format	One MCAP file per episode, with embedded Protobuf schemas
Perspective	Head-mounted egocentric recording of human bimanual activity
Video	Six synchronized H.264 camera streams
Motion	Head and stereo-camera poses; three-dimensional keypoints for both hands
Other sensors	Camera intrinsics and extrinsics; accelerometer, gyroscope, and IMU calibration
Language labels	English episode-level task description and time-bounded subtask descriptions

Data Modalities

Multi-view video

The camera streams are stereo_left, stereo_right, slam_left, slam_right, body_left, and body_right. The forward-facing stereo pair is 1920 × 1200; the other cameras are 640 × 480. Video messages contain H.264 frames in Annex-B format. The recording specification is 60 frames per second.

Each camera also has intrinsic calibration and a pose in the shared world coordinate frame. Intrinsics include the camera matrix and distortion parameters; the stereo images are undistorted, while the fisheye cameras use an equidistant distortion model.

Poses, hands, and IMU

/pose/head, /pose/stereo_left_cam, and /pose/stereo_right_cam provide positions and orientations. /pose/left_hand and /pose/right_hand provide ordered three-dimensional keypoints: wrist, thumb, index, middle, ring, and pinky joints through the fingertips. A hand may have no message for a frame when it is not detected.

The accelerometer and gyroscope streams contain timestamped three-axis measurements. An IMU calibration message provides biases, scale factors, noise parameters, and timing information. Distances are in meters; angular velocity is in radians per second and acceleration in meters per second squared. Poses and hand keypoints use the device SLAM world frame.

Annotations and Metadata

EgoSight provides natural-language supervision at two levels:

Annotation	Location	Meaning
Whole-task description	/session/metadata → task_info.task_description	An English description of the activity performed in the episode.
Subtask description	/annotation/semantic_segments → segment.subtask_description	An English description of a temporally localized step.
Subtask boundaries	segment.start_time, segment.end_time	Start and end in seconds relative to the first video frame.

The episode metadata also includes the scene, episode and task identifiers, anonymized operator identifier, recording-device information, and collection times. These fields support indexing and grouping; they are distinct from the language labels.

For example, an episode about cleaning a small box can be described as a whole task, while separate segments describe taking the box out, brushing its surfaces, and putting it back. Segment descriptions can name the hand, object, and action involved. They are time-bounded text annotations, not per-frame action classes. The MCAP schema contains an optional segment.skill list, but no populated skill taxonomy is specified here.

The released schema has no dedicated topics for audio, object bounding boxes, robot actions, rewards, or episode-success labels. Downstream users should not infer those labels from the available task descriptions.

Directory Layout

<scene>/
  <episode_uuid>.mcap

• <scene> is the scene name in English, for example Factory or Office.
• Each .mcap file holds one complete episode. The file name is the episode's unique ID.

MCAP is a general-purpose container format for multi-sensor recordings. You can open the files directly in a visualization tool or read them from Python (see the end of this document). All messages are Protobuf-encoded, and their schemas are embedded in each file, so you do not need to download anything else.

Conventions

• Time: all timestamps are Unix time in nanoseconds.
• Units: lengths are in meters, angular velocity is in rad/s and acceleration is in m/s².
• Quaternions: all quaternions are unit quaternions with w ≥ 0. Field order varies between message types, so always access the components by name (w, x, y, z).
• Coordinate frame: all poses and hand keypoints are expressed in a single world frame, world, which is defined by the device's SLAM system.

Topics

Topic	Content	Rate
/session/metadata	Episode metadata	1 message
/annotation/semantic_segments	Task segment annotations	1 message per segment
/sensor/camera/<camera>/video	H.264 video frames	1 message per frame
/sensor/camera/<camera>/intrinsic	Camera intrinsics	1 message per frame
/sensor/camera/<camera>/extrinsic	Camera pose in the world frame	1 message per frame
/pose/head	Head pose	1 message per frame
/pose/stereo_left_cam	Pose of the left stereo camera (stereo_left)	1 message per frame
/pose/stereo_right_cam	Pose of the right stereo camera (stereo_right)	1 message per frame
/pose/left_hand	21 left-hand keypoints	1 message per frame
/pose/right_hand	21 right-hand keypoints	1 message per frame
/sensor/imu/accel	Accelerometer	~400 Hz
/sensor/imu/gyro	Gyroscope	~400 Hz
/sensor/imu/calibration	IMU calibration parameters	1 message

<camera> is one of the following 6 cameras:

Camera	Resolution	Description
stereo_left	1920×1200	Forward-facing stereo pair, left camera, undistorted
stereo_right	1920×1200	Forward-facing stereo pair, right camera, undistorted
slam_left	640×480	Fisheye camera
slam_right	640×480	Fisheye camera
body_left	640×480	Fisheye camera
body_right	640×480	Fisheye camera

Message Fields

Metadata: /session/metadata

Field	Description
task_info.episode_uuid	Unique ID of the episode; matches the file name
task_info.episode_id	Original ID on the collection platform
task_info.scene	Scene name in English
task_info.task_description	English description of the whole task
task_info.task_id	Collection task ID
operator.device_id	Recording device ID
operator.operator_uuid	Anonymized operator ID
devices[]	Resolution, frame rate and start/end time (milliseconds) of each camera
global_time	Start and end time of the episode (UTC)

Task Segments: /annotation/semantic_segments

Each episode is divided into consecutive subtasks. There is one message per subtask:

Field	Description
task_description	Description of the whole task
segment.subtask_description	English description of this segment
segment.start_time	Start time of this segment, in seconds from the first video frame
segment.end_time	End time of this segment, in seconds from the first video frame

Video: /sensor/camera/<camera>/video

Video frames use the Foxglove CompressedVideo message. Each data field holds one H.264 frame in Annex-B format. Keyframes include SPS/PPS, so you can pass them straight to a decoder.

Camera Intrinsics and Extrinsics

• Intrinsics use the Foxglove CameraCalibration message. K is the 3×3 intrinsic matrix.
  • Fisheye cameras use the equidistant distortion model, with 4 distortion coefficients in D.
  • The stereo cameras are already undistorted, so D is empty.
• Extrinsics use the Foxglove FrameTransforms message. They give the camera pose in the world frame (from world to <camera>_camera).

Head and Camera Poses

/pose/head, /pose/stereo_left_cam and /pose/stereo_right_cam share the same format. Each message contains one entry in transforms, with a position pos (x, y, z) and an orientation quat (w, x, y, z), both in the world frame.

Hand Keypoints

/pose/left_hand and /pose/right_hand share the same format. Each message contains 21 entries in transforms, one per keypoint, in the order shown below. Each entry gives the keypoint position in the world frame. Its quat is always the identity quaternion and carries no information.

0      Wrist
1-4    Thumb_1  ~ Thumb_4
5-8    Index_1  ~ Index_4
9-12   Middle_1 ~ Middle_4
13-16  Ring_1   ~ Ring_4
17-20  Pinky_1  ~ Pinky_4

Index 1 is closest to the palm and index 4 is the fingertip. If a hand is not detected in a frame, that frame has no message for that hand, so a hand topic may contain fewer messages than there are video frames.

IMU: /sensor/imu/*

• /sensor/imu/accel and /sensor/imu/gyro: each message is one sample with x, y, z. The unit is given in the unit field.
• /sensor/imu/calibration: IMU biases, scale factors and noise parameters. The raw_json field contains the original calibration file.

Reading the Data

Install the dependencies:

pip install mcap mcap-protobuf-support

from mcap.reader import make_reader
from mcap_protobuf.decoder import DecoderFactory

with open("xxx.mcap", "rb") as f:
    reader = make_reader(f, decoder_factories=[DecoderFactory()])
    for schema, channel, message, msg in reader.iter_decoded_messages(
        topics=["/session/metadata", "/pose/left_hand"]
    ):
        print(channel.topic, message.log_time)
        if channel.topic == "/pose/left_hand":
            points = [(t.pos.x, t.pos.y, t.pos.z) for t in msg.transforms]  # 21 keypoints

License

EgoSight-Sample-500H is released under Creative Commons Attribution 4.0 International. Users must provide appropriate attribution and indicate whether changes were made.