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

RTK-SLAM Absolute Accuracy → FiftyOne (Native Multimodal MCAP)#

preview

The RTK-SLAM dataset from the Institute for Photogrammetry and Geoinformatics, University of Stuttgart, published as Willyzw/rtk-slam-dataset, converted to native multimodal MCAP episodes.

A handheld rig carrying a Livox MID360, a global shutter camera and a GNSS receiver is walked through a public park and a construction hall. Both block the sky in places, including a 30 m underpass and an indoor hall spanning more than 400 seconds of the route.

The reference is not the GNSS. A geodetic total station surveyed checkpoints along each route independently, so absolute error can be measured without first fitting the estimate onto the reference. The authors’ point is that the usual SE(3)-aligned error absorbs global drift and can understate the real figure by up to 76%.

Installation#

pip install fiftyone

Usage#

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

dataset = fouh.load_from_hub(
    "Voxel51/RTK-SLAM-Absolute-Accuracy",
    name="RTK-SLAM-Absolute-Accuracy",
    persistent=True,
)
fo.launch_app(dataset)

Where the sky was hardest to see:

view = dataset.sort_by("gnss_fix_rate")

What you get#

Four sequences and 64.6 minutes of walking. Each episode carries:

  • /camera, the handheld view at 1600x1200 and 20 Hz, as foxglove.CompressedVideo

  • /camera-calibration, the pinhole intrinsics and radtan coefficients

  • /lidar, the Livox MID360 sweep at 10 Hz, as foxglove.PointCloud

  • /imu.plot, accelerometer and gyroscope at 200 Hz

  • /gnss, the receiver’s fix, as foxglove.LocationFix

  • /gnss-quality.plot, the reported horizontal sigma and whether a fix was held at all

  • /checkpoints, the surveyed points as labelled spheres, coloured by whether each was taken outdoors, indoors or in the transition between

  • /pose-fast-lio-sam, /pose-okvis-lvig and /pose-okvis-vig, the three published trajectories

  • /trajectories, the same three as line strips

  • /checkpoint-error.plot, each system’s distance to each checkpoint

  • /tf, the rig geometry from the calibration

  • /instruction, the route being walked

Across the whole set that comes to 76,409 camera frames, 38,175 LiDAR sweeps, 763,673 IMU samples, 38,202 GNSS fixes and 87 surveyed checkpoints.

Episodes carry the fields sequence, site, environments, duration, num_checkpoints, num_camera_frames, num_lidar_sweeps, num_imu_samples, num_gnss_fixes, gnss_fix_rate, best_system, and error_<system>_rmse_m and error_<system>_max_m for each of the three.

Sequence

Duration

Checkpoints

GNSS fix

fast_lio_sam

okvis_lvig

okvis_vig

stadtgarten_seq1

26m43s

36

54%

0.16 m

4.11 m

3.28 m

stadtgarten_seq2

15m35s

19

40%

0.17 m

3.19 m

2.69 m

construction_seq1

12m21s

16

25%

0.26 m

0.75 m

1.44 m

construction_seq2

9m60s

16

23%

0.46 m

0.81 m

3.72 m

Errors are RMSE over that sequence’s checkpoints, with no alignment applied.

Notes on the conversion#

The release publishes the same recordings three times, as ROS1 bags, ROS2 stores and an extended EuRoC layout. The ROS2 store is the one converted.

The surveyed checkpoints are in ETRS89 / UTM 32N with DHHN2016 heights, which are metres above the geoid, while the recordings and the published trajectories use ellipsoidal height. Around Stuttgart the two differ by about 48 m, so the checkpoints are transformed through the geoid model before being placed in the episode’s local east-north-up frame. That frame is anchored on the origin the trajectory authors used, so the estimates need no adjustment.

The GNSS is carried as recorded. It is an input to the systems under test rather than a reference, and it is degraded on purpose: on the construction sequences a fix is held for under a quarter of the run.

Video is re-encoded to Annex-B H.264 without B-frames. The source carries JPEG frames, which are most of its volume.

The /livox/lidar topic duplicates /livox/points in a vendor-specific message type and is not carried.

stadtgarten_seq2 runs 935 s in the recording against the 14m36s its source table gives. The other three match their published durations to the second.

License & attribution#

The source release is distributed under CC BY 4.0, and this conversion is distributed under the same license.

@article{zhang2025rtkslam,
  title   = {An RTK-SLAM Dataset for Absolute Accuracy Evaluation in
             GNSS-Degraded Environments},
  author  = {Zhang, Wei and Ress, Vincent and Skuddis, David and
             Soergel, Uwe and Haala, Norbert},
  journal = {ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial
             Information Sciences},
  year    = {2025},
}

Changes from the source: conversion to the FiftyOne MCAP flavor, re-encoding of the video to H.264, transformation of the surveyed checkpoints into the trajectory frame, and encoding of the sensor, pose and error streams as message streams.