Note

This is a Hugging Face dataset. For large datasets, ensure huggingface_hub>=1.1.3 to avoid rate limits. Learn more in the Hugging Face integration docs.

Hugging Face

Dataset Card for Edged-USLAM DAVIS346 drone event camera#

preview

This is a FiftyOne dataset with 13 samples. Each sample is one flight of a quadrotor carrying an event camera, stored as a native multimodal MCAP episode.

The source is the Edged-USLAM dataset, recorded for Edged USLAM: Edge-Aware Event-Based SLAM with Learning-Based Depth Priors (ICRA 2026), converted from ROS 1 bags. A DAVIS346 event camera rides a quadrotor flown in a motion-capture room. Each run records the camera’s asynchronous events, its grayscale frames and its inertial unit, together with the Vicon pose of the vehicle. The motion runs fly lines, squares, circles, aggressive turns and manual flights, some through an IR-filtered lens and some through the default lens, and the illumination runs fly under lighting from under 5 lux to blinking lights and strong side light.

Installation#

If you haven’t already, install FiftyOne:

pip install -U fiftyone

Usage#

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

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = fouh.load_from_hub(
    "Voxel51/Edged-USLAM-Event-Camera",
    name="Edged-USLAM-Event-Camera",
    persistent=True,
)

# Launch the App
session = fo.launch_app(dataset)

Dataset Details#

Dataset Description#

13 runs, 5 illumination and 8 motion, totalling 16m 27s of flight, 657,511,373 events and 39,457 frames.

  • Curated by: Şebnem Sarıözkan, Hürkan Şahin, Olaya Álvarez-Tuñón and Erdal Kayacan (source release)

  • Funded by: [More Information Needed]

  • Shared by: Voxel51 (FiftyOne conversion)

  • Language(s): Not applicable (sensor data)

  • License: CC BY 4.0

Dataset Sources#

Uses#

Direct Use#

Event-based SLAM and visual-inertial odometry from a DAVIS346 event camera, its grayscale frames and its inertial unit, evaluated against the Vicon pose of the vehicle shipped with each run, across different motion profiles and lighting conditions.

Out-of-Scope Use#

[More Information Needed]

Dataset Structure#

Topology

An ungrouped FiftyOne dataset with media_type="multimodal". One sample is one run, and its filepath is a .fo.mcap file stored under data/illumination/ or data/motion/. There are 13 samples (5 in the illumination group, 8 in the motion group), no sample tags, and an empty dataset.info. The dataset has no FiftyOne label fields; every signal lives inside the MCAP file as a channel and is shown by the App’s multimodal viewer.

Sample fields

Field

FiftyOne type

Description

id, filepath, tags, metadata, created_at, last_modified_at

built-in

Standard FiftyOne sample fields

sequence

StringField

Run name, e.g. 30_lit (derived from the source bag)

group

StringField

Run group: illumination or motion

lighting

StringField

Lighting condition of an illumination run (e.g. under 5 lux, blinking lights); unset for motion runs

recorded

StringField

Recording date

duration

FloatField

Run length in seconds

num_events

IntField

Events the camera produced

num_frames

IntField

Grayscale frames

num_imu_samples

IntField

Samples from the camera’s inertial unit

num_ground_truth_poses

IntField

Poses on the Vicon ground-truth trajectory

mean_event_rate_mev_s

FloatField

Mean event rate over the run, in millions of events per second

peak_event_rate_mev_s

FloatField

Event rate of the busiest 1/30 s window, in millions of events per second

mean_frame_brightness

FloatField

Mean pixel value of the grayscale frames

ground_truth_path_m

FloatField

Length of the Vicon ground-truth trajectory in metres

lens

StringField

filtered or unfiltered, set for the runs the release files under filtered/ and unfiltered/ (5 runs); unset for the other 8

Episode contents

Each MCAP episode contains these channels:

Channel

Schema or content

/events

Every event the camera produced, in windows of 1/30 s, as foxglove.PointCloud with x and y the 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

A render of each window at 346x260, ON events white and OFF events black on gray, foxglove.CompressedVideo

/frames

The grayscale frames at 346x260 and 40 Hz, foxglove.CompressedVideo

<video channel>-calibration

A calibration topic beside each video, foxglove.CameraCalibration

/tf

The camera’s pose in the inertial unit’s frame, foxglove.FrameTransform

/imu.plot

The camera’s inertial unit at 1000 Hz

/ground-truth, /ground-truth.plot

The Vicon pose of the vehicle as foxglove.PoseInFrame, with its position as a plot

/sequence

Names the run

Runs

Run

Group

Lens

Lighting

Duration

Events

Frames

Mean event rate

Path

30_lit

illumination

30% lighting

1m 01s

47,361,052

2,457

0.77 M/s

39.9 m

60_lit

illumination

60% lighting

1m 10s

55,798,203

2,776

0.80 M/s

36.5 m

dyn_HDR

illumination

blinking lights

1m 11s

56,369,994

2,820

0.80 M/s

35.3 m

low_lit

illumination

under 5 lux

1m 06s

34,494,687

2,621

0.53 M/s

38.0 m

static_HDR

illumination

strong side light

1m 08s

46,852,643

2,713

0.69 M/s

35.1 m

auto_circle_unfiltered

motion

unfiltered

0m 50s

39,491,965

2,021

0.78 M/s

41.6 m

circle_filtered

motion

filtered

0m 41s

27,209,124

1,641

0.66 M/s

38.1 m

dark_manual_filtered

motion

filtered

0m 50s

19,953,129

1,989

0.40 M/s

29.8 m

line

motion

1m 23s

26,434,471

3,317

0.32 M/s

19.1 m

manual_long_unfiltered_x

motion

unfiltered

1m 56s

121,465,644

4,647

1.04 M/s

46.7 m

manual_unfiltered_y

motion

unfiltered

1m 15s

70,610,958

3,007

0.94 M/s

52.6 m

square

motion

1m 17s

49,516,705

3,099

0.64 M/s

49.4 m

turner

motion

2m 39s

61,952,798

6,349

0.39 M/s

42.1 m

Parsing decisions

  • Why MCAP: the dataset is one multimodal sample per run, so the events, grayscale frames, inertial unit and ground-truth pose play back on one shared clock instead of being split into per-frame samples.

  • ROS 1 bags: the bags were read without a ROS install.

  • Events: every event in every event array is carried, on the camera’s own microsecond timestamps. The windows run from the first event of each run, so a window’s events and its render share one stamp, and an event’s time is that stamp less 1/30 s plus its z in milliseconds. x and y are the pixel’s column and row, so a 3D panel set to the events frame shows each window as a volume of pixel position against time.

  • Video encoding: the grayscale frames are re-encoded to Annex-B H.264 without B-frames, one access unit per frame, on each frame’s own timestamp.

  • Calibration: the calibration is the release’s own DAVIS_346.yaml: a pinhole camera with radial-tangential distortion, carried as plumb_bob, which the release gives for both the events and the frames. Its T_B_C, the camera’s pose in the body frame of the inertial unit, is carried as /tf.

  • Ground truth: the release records the Vicon ground truth on /mavros/vision_pose/pose for the illumination runs and the line, square and turner runs, and on /local_pose_vicon/pose for the filtered and unfiltered runs. Both are carried as /ground-truth, in the frame each names, camera_odom_frame and world.

  • Inertial unit: it is carried at the rate the bags record it.

  • Recording dates: the recorded field holds 2025-04-14, 2025-05-02 or 2025-07-16 for 12 runs; auto_circle_unfiltered holds 1970-01-01.

Dataset Creation#

Curation Rationale#

[More Information Needed]

Source Data#

Data Collection and Processing#

Recorded with a DAVIS346 event camera on a quadrotor flown in a motion-capture room, with the vehicle’s pose recorded by Vicon. The FiftyOne conversion reads the ROS 1 bags without a ROS install and writes one MCAP episode per bag; see Parsing decisions above for the changes made.

Who are the source data producers?#

The authors of Edged USLAM: Edge-Aware Event-Based SLAM with Learning-Based Depth Priors.

Annotations#

Annotation process#

The dataset has no human annotations. The ground-truth trajectory in each run is the Vicon pose of the vehicle as recorded by the source release.

Who are the annotators?#

Not applicable.

Personal and Sensitive Information#

[More Information Needed]

Citation#

The authors ask that work using the dataset cite:

BibTeX:

@inproceedings{sariozkan2026edged,
  title={Edged USLAM: Edge-Aware Event-Based SLAM with Learning-Based Depth Priors},
  author={Sarıözkan, Şebnem and Şahin, Hürkan and Álvarez-Tuñón, Olaya and Kayacan, Erdal},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026}
}

APA:

Sarıözkan, Ş., Şahin, H., Álvarez-Tuñón, O., & Kayacan, E. (2026). Edged USLAM: Edge-aware event-based SLAM with learning-based depth priors. In IEEE International Conference on Robotics and Automation (ICRA).

More Information#

The source release is distributed under the Creative Commons Attribution 4.0 International license (CC-BY-4.0), and this conversion is distributed under the same license.

Changes from the source: conversion from ROS 1 bags to the FiftyOne MCAP flavor, the events cut into 1/30 s windows carried as point clouds with a grayscale render of each window, H.264 encoding of the grayscale frames, the release’s calibration carried as camera intrinsics and a transform, and the two ground-truth topics carried as one.

Dataset Card Authors#

[More Information Needed]

Dataset Card Contact#

[More Information Needed]