Voxel51/Edged-USLAM-Event-Camera

Edged-USLAM DAVIS346 drone event camera → FiftyOne (Native Multimodal MCAP)

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 to native multimodal MCAP episodes.

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 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.

13 runs, 16m 27s of flight, 657,511,373 events and 39,457 frames.

Installation

pip install fiftyone

Usage

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

dataset = fouh.load_from_hub(
    "Voxel51/Edged-USLAM-Event-Camera",
    name="Edged-USLAM-Event-Camera",
    persistent=True,
)
fo.launch_app(dataset)

The darkest runs, where the frames see least and the events still see the room:

view = dataset.sort_by("mean_frame_brightness")
fo.launch_app(dataset, view=view)

What you get

Each episode contains:

• /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, as foxglove.CompressedVideo
• /frames, the grayscale frames at 346x260 and 40 Hz, as foxglove.CompressedVideo
• a -calibration topic beside each video, as foxglove.CameraCalibration
• /tf, the camera's pose in the inertial unit's frame, as foxglove.FrameTransform
• /imu.plot, the camera's inertial unit at 1000 Hz
• /ground-truth, the Vicon pose of the vehicle, as foxglove.PoseInFrame, with its position on /ground-truth.plot
• /sequence, naming the run

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

Episodes carry the fields sequence, group, lens, lighting, recorded, duration, num_events, num_frames, num_imu_samples, num_ground_truth_poses, mean_event_rate_mev_s, peak_event_rate_mev_s, mean_frame_brightness and ground_truth_path_m. lens is set for the runs the release files under filtered/ and unfiltered/, and lighting for the illumination runs; peak_event_rate_mev_s is the busiest 1/30 s window, and mean_frame_brightness the mean pixel value of the grayscale frames.

Notes on the conversion

The bags are ROS 1 and were read without a ROS install. 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.

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. 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, and its T_B_C, the camera's pose in the body frame of the inertial unit, carried as /tf.

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. The inertial unit is carried at the rate the bags record it.

License & attribution

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. The authors ask that work using the dataset cite:

@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}
}

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.