Voxel51/ColoRadar-Sample

ColoRadar sample → FiftyOne (Native Multimodal MCAP)

7 sequences from ColoRadar, the 3D millimetre-wave radar dataset from the Autonomous Robotics and Perception Group at the University of Colorado Boulder, converted from ROS 1 bags to native multimodal MCAP episodes.

A handheld rig carries two FMCW radars, a TI MMWCAS-RF-EVM cascaded imaging radar and a TI AWR1843BOOST single-chip radar, beside an Ouster OS1-64 LiDAR and a Lord Microstrain 3DM-GX5-25 inertial unit. The release holds 52 sequences recorded in seven places: hallways, a lab, a large motion-capture space, outdoor built environments, the narrow and the large passages of an underground mine, and a fast ride along paths and roads. This sample carries one sequence from each place, the shortest of its group.

13m 46s of recording, 8,263 LiDAR scans holding 382.6 million points, 4,145 cascaded radar frames and 8,340 single-chip radar scans.

Installation

pip install fiftyone

Usage

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

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

The sequences that covered the most ground:

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

What you get

Each episode contains:

• /lidar-points, the Ouster OS1-64 scans, as foxglove.PointCloud with x, y, z, intensity, reflectivity and ring
• /single-chip-radar-points, the AWR1843's own point clouds, as foxglove.PointCloud with x, y, z, intensity, range and doppler
• /cascade-radar-points, points derived from each cascaded radar heatmap, as foxglove.PointCloud with x, y, z, intensity and doppler
• /cascade-radar-bev, a top-down render of each cascaded radar heatmap, as foxglove.CompressedVideo
• /imu.plot, the Microstrain inertial unit at about 100 Hz
• /ground-truth, the release's LiDAR-inertial trajectory, as foxglove.PoseInFrame, with its position on /ground-truth.plot
• /motion-capture, the Vicon pose of the rig in the motion-capture space, with its position on /motion-capture.plot
• /tf, each sensor's pose in the rig's base frame, as foxglove.FrameTransform
• /sequence, naming the sequence and its environment

Sequence	Environment	Recorded	Duration	LiDAR scans	Cascade frames	Single-chip scans	Path
arpg_lab_run2	Indoor built environment	2020-12-21	1m 41s	1,014	504	1,015	67 m
aspen_run9	Large motion capture space	2021-02-24	1m 16s	765	414	836	42 m
ec_hallways_run4	Indoor built environment	2020-12-21	1m 30s	899	447	902	103 m
edgar_army_run5	Large subterranean environment	2021-02-23	1m 54s	1,141	567	1,141	108 m
edgar_classroom_run4	Narrow subterranean environment	2021-02-23	2m 44s	1,639	816	1,639	184 m
longboard_run1	High-speed multi-use paths and roads	2021-02-22	2m 57s	1,767	882	1,770	766 m
outdoors_run8	Outdoor built environment	2021-03-02	1m 44s	1,038	515	1,037	121 m

Episodes carry the fields sequence, environment, recorded, duration, num_lidar_scans, num_lidar_points, num_cascade_frames, num_cascade_points, num_single_chip_scans, num_single_chip_points, num_imu_samples, ground_truth_path_m and has_motion_capture. environment is the release's own description of each sequence's group, and ground_truth_path_m the length of the LiDAR-inertial trajectory.

Notes on the conversion

The bags are ROS 1 and were read without a ROS install. The LiDAR scans drop the returns the sensor marks empty, and the single-chip radar's point clouds are carried as recorded.

The cascaded radar publishes no point cloud, only a heatmap of 128 range by 128 azimuth by 32 elevation cells, each holding an intensity and a Doppler velocity. Its points are derived the way the release's plotting tool derives them: every cell beyond the tenth range bin whose intensity exceeds a fifth of the way from the frame's weakest cell to its strongest, placed at the cell's centre. The release's tool also thins them on a 0.3 m grid; these are not thinned. /cascade-radar-bev draws each heatmap from above, the strongest cell in each column of elevations on a scale of 30 dB below the frame's peak, with the radar facing up.

/tf is the release's calibration, each sensor's pose in the rig's base frame, with the single-chip radar's heatmap frame dca_link placed on its point-cloud frame as the bags place it. The raw ADC samples of both radars and the single-chip radar's heatmaps are not carried, nor are the bags' own /tf stream and the LiDAR's built-in inertial unit; all are in the release.

License & attribution

The source release is distributed under the Apache License 2.0, and this conversion is distributed under the same license, whose text is in LICENSE beside this card. Cite:

@article{kramer2022coloradar,
  title={ColoRadar: The direct 3D millimeter wave radar dataset},
  author={Kramer, Andrew and Harlow, Kyle and Williams, Christopher and Heckman, Christoffer},
  journal={The International Journal of Robotics Research},
  volume={41},
  number={4},
  pages={351--360},
  year={2022},
  doi={10.1177/02783649211068535}
}

Changes from the source: 7 of the release's 52 sequences, converted from ROS 1 bags to the FiftyOne MCAP flavor, the LiDAR's empty returns dropped, points derived from the cascaded radar's heatmaps with a top-down render of each, the release's calibration carried as transforms, and the raw ADC samples, the single-chip heatmaps and the bags' own /tf left out.