Voxel51/M3ED-Sample

M3ED sample → FiftyOne (Native Multimodal MCAP)

3 sequences from M3ED, the multi-robot, multi-sensor, multi-environment event dataset from the GRASP Laboratory at the University of Pennsylvania, converted to native multimodal MCAP episodes.

A car, a quadrotor and a Boston Dynamics Spot carry the same sensor head: a stereo pair of Prophesee EVK4 HD event cameras at 1280x720, a stereo pair of grayscale cameras and a color camera at 1280x800, an inertial unit and an Ouster OS1-64 LiDAR, through cities, forests and buildings by day and night. The release ships ground-truth poses and depth for the left event camera beside each sequence. This sample carries one sequence from each platform: the car, car_urban_day_horse; the quadrotor, falcon_outdoor_night_high_beams; Spot, spot_indoor_stairwell.

192 seconds of recording, 3,631,437,863 events, 4,802 image triplets and 1,921 LiDAR scans.

Installation

pip install fiftyone

Usage

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

dataset = fouh.load_from_hub(
    "Voxel51/M3ED-Sample",
    name="M3ED-Sample",
    persistent=True,
)
fo.launch_app(dataset)

The busiest event streams first:

view = dataset.sort_by("peak_event_rate_mev_s", reverse=True)
fo.launch_app(dataset, view=view)

What you get

Each episode contains:

• /events-left and /events-right, every event each event camera produced, in windows of 1/30 s, as foxglove.PointCloud with x and y the raw pixel, z the time since the window opened in milliseconds and polarity 1 for a brightness increase and 0 for a decrease; each message is stamped at its window's close
• /event-frames-left and /event-frames-right, a render of each window, ON events white and OFF events black on gray, as foxglove.CompressedVideo
• /images-left and /images-right, the grayscale pair, and /images-rgb, the color camera, at 1280x800, as foxglove.CompressedVideo
• /lidar-points, the Ouster scans, as foxglove.PointCloud with x, y, z, reflectivity and signal
• /imu.plot, the sensor head's inertial unit
• /ground-truth, the left event camera's pose relative to its first ground-truth pose, as foxglove.PoseInFrame, with its position on /ground-truth.plot
• /depth-ground-truth, the depth ground truth for the left event camera, as foxglove.CompressedImage (16-bit PNG, millimetres, 0 where there is none)
• a -calibration topic beside each camera stream, as foxglove.CameraCalibration
• /tf, each sensor's pose in the left event camera's frame, as foxglove.FrameTransform
• /sequence, naming the sequence and its platform

Sequence	Platform	Environment	Duration	Events	Peak event rate	LiDAR scans	Path
car_urban_day_horse	car	urban day	28.7 s	846,851,622	45.9 M/s	286	47 m
falcon_outdoor_night_high_beams	falcon	outdoor night	64.9 s	901,446,929	47.4 M/s	648	46 m
spot_indoor_stairwell	spot	indoor	98.9 s	1,883,139,312	63.8 M/s	987	30 m

Episodes carry the fields sequence, platform, environment, recorded, duration, num_events_left, num_events_right, peak_event_rate_mev_s, num_images, num_lidar_scans, num_lidar_points, num_ground_truth_poses, ground_truth_path_m, valid_depth_fraction and mean_image_brightness. peak_event_rate_mev_s is the busier camera's busiest 1/30 s window, valid_depth_fraction the share of depth ground-truth pixels holding a value, and mean_image_brightness the mean pixel value of the color camera.

Notes on the conversion

The release's timestamps run in microseconds from each recording's start, which its stats file gives in Unix time; every stream is placed on that clock. Every event in each event camera's data is carried, at the raw pixel, with each camera's distortion on its calibration topic. Both cameras' windows run from the first event of either, so an event's time is its window's stamp less 1/30 s plus its z in milliseconds.

The cameras are re-encoded to Annex-B H.264 without B-frames, one access unit per frame. The release stores the color camera's pixels in blue-green-red order, and they are carried in red-green-blue order, keeping the green cast they have in the release. The LiDAR packets the release stores are decoded to points with the Ouster SDK, keeping the returns with a range. The release gives the ground-truth pose as the transform from the first pose's frame into the current one, so /ground-truth carries its inverse, the camera's pose in the first pose's frame. The depth ground truth, which the release stores as 32-bit floats in metres, is carried as 16-bit PNG in millimetres, rounded. The release's T_to_prophesee_left for each sensor maps that sensor's points into the left event camera, which is the sensor's pose in that frame, and is carried as /tf. A few inertial samples the release marks as untimed are left out, as are the LiDAR's own inertial unit and the semantic labels.

License & attribution

M3ED is distributed under the Creative Commons Attribution-ShareAlike 4.0 International license (CC-BY-SA-4.0), and this conversion is distributed under the same license. Cite:

@InProceedings{Chaney_2023_CVPR,
  author = {Chaney, Kenneth and Cladera, Fernando and Wang, Ziyun and Bisulco, Anthony and Hsieh, M. Ani and Korpela, Christopher and Kumar, Vijay and Taylor, Camillo J. and Daniilidis, Kostas},
  title = {M3ED: Multi-Robot, Multi-Sensor, Multi-Environment Event Dataset},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  month = {June},
  year = {2023},
  pages = {4015-4022}
}

Changes from the source: 3 of the release's sequences, converted from HDF5 to the FiftyOne MCAP flavor, each event camera's events cut into 1/30 s windows carried as point clouds with a grayscale render of each window, H.264 encoding of the cameras with the color camera's channels in red-green-blue order, the LiDAR packets decoded to points, the ground-truth poses inverted, the depth ground truth carried as 16-bit millimetres, and untimed inertial samples, the LiDAR's inertial unit and the semantic labels left out.