Voxel51/DSEC-Sample

DSEC sample → FiftyOne (Native Multimodal MCAP)

6 sequences from DSEC, the stereo event camera dataset for driving from the Robotics and Perception Group at the University of Zurich, converted to native multimodal MCAP episodes.

A car carries a stereo pair of Prophesee event cameras at 640x480 and a stereo pair of global-shutter color cameras at 1440x1080 through Zurich, Thun and Interlaken. Disparity derived from LiDAR scans is the ground truth for both pairs, and some sequences add optical flow ground truth for the event cameras. The release holds 53 sequences; this sample carries 6 of the training sequences, which come with their ground truth, chosen across the three places.

105 seconds of driving, 2,233,378,133 events and 2,112 image pairs.

Installation

pip install fiftyone

Usage

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

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

The sequences with the busiest event streams:

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 rectified color images at 20 Hz, as foxglove.CompressedVideo
• /disparity-event and /disparity-image, the disparity ground truth for each stereo pair at 10 Hz, as foxglove.CompressedImage (16-bit PNG, 256 per pixel of disparity, 0 where there is none)
• /optical-flow-forward and /optical-flow-backward on the sequences that have them, the event cameras' optical flow ground truth, as foxglove.CompressedImage (16-bit PNG)
• a -calibration topic beside each camera stream, as foxglove.CameraCalibration
• /tf, the four cameras and their rectified frames, as foxglove.FrameTransform
• /sequence, naming the sequence and its place

Sequence	Place	Duration	Events	Peak event rate	Image pairs	Disparity maps	Flow maps
interlaken_00_c	Interlaken	26.8 s	884,252,854	28.8 M/s	537	269
thun_00_a	Thun	11.9 s	261,111,191	20.5 M/s	239	120	41
zurich_city_02_a	Zurich	11.7 s	312,356,446	22.3 M/s	235	118	64
zurich_city_04_b	Zurich	13.4 s	247,359,435	20.6 M/s	269	135
zurich_city_09_b	Zurich	18.3 s	298,543,673	22.6 M/s	367	184
zurich_city_11_a	Zurich	23.2 s	229,754,534	12.4 M/s	465	233	231

Episodes carry the fields sequence, place, duration, num_events_left, num_events_right, peak_event_rate_mev_s, num_images, num_disparity_maps, num_optical_flow_maps and mean_image_brightness. peak_event_rate_mev_s is the busier camera's busiest 1/30 s window, and mean_image_brightness the mean pixel value of the left color images.

Notes on the conversion

Every event in each event camera's file is carried, on the timestamps the release records in microseconds on the rig's own clock. Both cameras' windows run from the first event of either, so the two cameras share their window stamps, and an event's time is its window's stamp less 1/30 s plus its z in milliseconds. The events are carried at the raw, distorted pixel, with each event camera's distortion on its calibration topic.

The color images are the release's rectified images, re-encoded to Annex-B H.264 without B-frames, one access unit per frame, on the release's image timestamps. The disparity and optical flow maps are carried as the release ships them. /disparity-event lies in the rectified frame of the left event camera and /disparity-image in that of the left color camera; each optical flow map is stamped at the start of the interval it covers. The optical flow PNGs hold three 16-bit channels: horizontal and vertical flow as (value - 2^15) / 128 pixels, and a validity flag.

/tf is the release's calibration. The release maps points from each camera into the next and from each camera into its rectified frame; each transform is carried as the inverse, the child's pose in its parent. The raw LiDAR and the semantic labels the release offers are not carried.

License & attribution

DSEC 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:

@Article{Gehrig21ral,
  author = {Mathias Gehrig and Willem Aarents and Daniel Gehrig and Davide Scaramuzza},
  title = {DSEC: A Stereo Event Camera Dataset for Driving Scenarios},
  journal = {IEEE Robotics and Automation Letters},
  year = {2021},
  doi = {10.1109/LRA.2021.3068942}
}

@InProceedings{Gehrig3dv2021,
  author = {Mathias Gehrig and Mario Millh\"ausler and Daniel Gehrig and Davide Scaramuzza},
  title = {E-RAFT: Dense Optical Flow from Event Cameras},
  booktitle = {International Conference on 3D Vision (3DV)},
  year = {2021}
}

Changes from the source: 6 of the release's training sequences, converted 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 rectified color images, the release's calibration carried as camera intrinsics and inverted transforms, and the raw LiDAR and semantic labels left out.