The Dataset Viewer has been disabled on this dataset.

Elevator-VIGS Dataset

Handheld and simulated sequences for visual-inertial SLAM through elevator rides, released with Elevator-VIGS: Separating Elevator Motion from Robot Motion in Visual-Inertial Gaussian Splatting SLAM (paper, code, project page).

  • 19 real-world sequences (70 min) in 10 buildings, recorded with a handheld Manifold Odin 1 (RGB 1600×1296 at 10 Hz, IMU 400 Hz, LiDAR at 10 Hz). The rig rests on the floor at a control point before and after the ride, and on the one-way sequences the height H between the two control points is laser-measured. MT-Pose.txt is the rig's own LiDAR-inertial estimate, which drifts through a ride and is not a ground truth.
  • 16 simulated sequences (36 min) in 4 buildings, rendered in NVIDIA Isaac Sim (RGB 640×480 at 20 Hz, IMU 400 Hz, OS1-32 LiDAR at 10 Hz). MT-Pose.txt is the ground truth.

Eight of the real-world sequences

Ten of the simulated sequences

Sequences

Real-world Itinerary Floors H [m] Elevator
Office1 one-way, up 4 14.27 enclosed
Office2 one-way, up 5 16.78 glass
Office3 round trip, up and back by elevator 2 – enclosed and glass
Office4 round trip, up by elevator, back by stairs 4 – enclosed
Office5 one-way, up 4 14.28 enclosed
Office6 one-way, up, via an intermediate floor 4 14.28 enclosed
Station1 one-way, down 1 7.29 glass
Station2 round trip, down by elevator, back by ramp 1 – glass
Station3 round trip, up by elevator, back by stairs 1 – glass
Residential1 one-way, up 5 19.79 enclosed
Residential2 one-way, up 15 49.98 enclosed
Residential3 round trip, up and back by elevator 2 – enclosed
Residential4 one-way, up, via an intermediate floor 15 49.97 enclosed
Campus1 one-way, up 2 13.65 enclosed
Campus2 one-way, up 4 15.07 enclosed
Campus3 one-way, up 5 15.97 enclosed
Campus4 round trip, down by elevator, back by stairs 1 – enclosed
Campus5 round trip, up by elevator, back by stairs 1 – enclosed
Campus6 round trip, up and back by elevator 6 – enclosed

Each simulated building has four sequences, -E (1 floor, one-way), -M (4 floors up in one elevator and back down in a second), -H (10 floors, one-way) and -Walk (no ride).

Simulated Elevator Rise of -E / -M / -H [m]
Office1 enclosed 3.35 / 13.40 / 33.50
Office2 enclosed 3.30 / 13.20 / 33.00
Mall glass 4.50 / 18.00 / 45.00
Factory enclosed 5.00 / 20.00 / 50.00

Download

Each sequence is two zips, <Seq>.zip (camera, IMU, reference trajectory, labels) and <Seq>_lidar.zip. Camera and IMU take 44 GB, everything 81 GB. manifest.csv lists frames, duration and checksums per sequence.

pip install -U huggingface_hub
hf download Rui5125/Elevator-VIGS --repo-type dataset --local-dir Elevator-VIGS                                               # everything
hf download Rui5125/Elevator-VIGS --repo-type dataset --local-dir Elevator-VIGS --include "*/*.zip" --exclude "*_lidar.zip"   # camera + IMU only
cd Elevator-VIGS && for z in */*.zip; do unzip -q -u -o "$z"; done   # -> Elevator-VIGS/{real,sim}/<Seq>/

Update, November 2026. Office5, Office6 and Residential4 are new, the four Mall sequences are re-rendered with a changed elevator cabin (same routes, trajectories and IMU), and the rides entries of Office4 and Station3 are corrected. Running the lines above again updates a copy of the first release, which remains at revision 188c830.

Format

real|sim/<Seq>/
  images/<stamp>.jpg|png     real: <seconds>.jpg; sim: <nanoseconds>.png
  imu.txt                    t,gx,gy,gz,ax,ay,az [rad/s, m/s^2]; real: seconds, one header line; sim: nanoseconds
  calib.txt                  fx fy cx cy, pinhole without distortion
  MT-Pose.txt                t x y z qw qx qy qz, pose of the IMU frame, t in seconds
  lidar/<stamp>.npz          data (N,3) xyz in the LiDAR frame, intensity (N,), timestamp (N,) in ns per point
  control_points.json        real only, see Evaluation
  events.json                sim only, see Evaluation
  MT-Cloud.ply, MT-Traj.ply  real only, the rig's LiDAR map and trajectory

All stamps of a sequence share one clock. T_cb maps IMU-frame points into the camera frame, and T_bl maps LiDAR-frame points into the IMU frame.

real  T_cb = [[ 2.38e-03, -1.0,       8.30e-04,  5.68021088e-02],
              [-1.00e-05, -8.30e-04, -1.0,       2.70171357e-02],
              [ 1.0,       2.38e-03, -1.00e-05,  2.45616807e-02],
              [ 0, 0, 0, 1]]
      T_bl = [[1, 0, 0, -0.0254], [0, 1, 0, 0.0330], [0, 0, 1, 0.0221], [0, 0, 0, 1]]
sim   T_cb = [[0, -1, 0, 0.00], [0, 0, -1, 0.02], [1, 0, 0, -0.03], [0, 0, 0, 1]]
      T_bl = [[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0.1], [0, 0, 0, 1]]

On both sets, the paper sets the IMU noise density and random walk to 2.08e-3 m/s²/√Hz and 4.13e-4 for the accelerometer and to 2.05e-4 rad/s/√Hz and 1.11e-5 for the gyroscope.

Evaluation

Real-world. control_points.json gives the window [t0, t1] in seconds of each rest at a control point, B1 on the departure floor, C1 on the arrival floor (one-way) and B2 back on the departure floor (round trip). Take the median height of the estimate in each window, along its own gravity direction, so no trajectory alignment is needed. The error is the B1→C1 rise minus H on a one-way sequence and the B1→B2 height difference on a round trip. rides lists an interval that contains each ride, read from the IMU.

Simulated. Compare the estimate with MT-Pose.txt. events.json lists every ride under legs, with its floors, rise and start and stop stamps.

The scorers of the paper are scripts/elevator_eval_utils.py and eval_elevator_mono.py in the code repository.

ROS bag

to_rosbag.py (needs numpy and rosbags, and opencv-python for --images) builds a ROS1 bag with /livox/lidar and /livox/imu from a sequence folder. --images adds /camera/image_raw, and --time-unit us writes the per-point time in microseconds instead of seconds.

python to_rosbag.py Elevator-VIGS/real/Campus1 --images --time-unit us

Privacy and license

Faces are blurred in every released real-world frame with deface, and there is no audio. The dataset is released under CC BY-NC 4.0. The shop furnishing of the simulated Mall comes from an asset pack under the same license.

Citation

@misc{zhou2026elevatorvigsseparatingelevatormotion,
      title={Elevator-VIGS: Separating Elevator Motion from Robot Motion in Visual-Inertial Gaussian Splatting SLAM},
      author={Rui Zhou and Zihan Zhu and Wei Zhang and Zizhou Luo and Norbert Haala and Marc Pollefeys},
      year={2026},
      eprint={2609.23491},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2609.23491},
}
Downloads last month
505

Paper for Rui5125/Elevator-VIGS