Elevator-VIGS Dataset

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

• 16 real-world sequences, handheld Manifold Odin 1 (RGB 1600×1296 at 10 Hz, IMU 400 Hz, LiDAR 120°×90° at 10 Hz) in offices, railway stations, a residential building and campus buildings. Rides span 1–15 floors and 7–50 m. The rig rests on the floor at a control point before and after each ride; on one-way sequences the floor-to-floor rise H between the two control points is laser-measured (±1.5 mm). MT-Pose.txt is the rig's own LiDAR-inertial solution: accurate on the walks, drifts through and after a ride, not a ground truth.
• 16 simulated sequences, NVIDIA Isaac Sim (RGB 640×480 at 20 Hz, IMU 400 Hz with noise and bias, Ouster OS1-32 LiDAR at 10 Hz) in four buildings, each with an easy (1 floor), medium (round trip, 4 floors, two elevators) and hard (10 floors) ride plus one elevator-free walk. MT-Pose.txt is the exact ground truth.

Sequences are named as in the paper and share one layout, real/<Seq>/ and sim/<Seq>/, so the same code runs on either arm.

Sequences

H is the laser-measured (real) or simulator (sim) rise of the one-way ride.

Real-world

Eight of the 16 sequences, at most one per building. Every tile is the LiDAR map the rig builds of that sequence, with its MT-Pose.txt trajectory and a pin at each control point. A one-way sequence carries two pins, the departure and the arrival floor; a round trip carries one, because the return is scored at the same control point.

Sequence	Itinerary	Floors	H [m]	Elevator	Frames	Duration [s]
Office1	one-way, up	4	14.27	enclosed	1393	136
Office2	one-way, up	5	16.78	glass	978	96
Office3	round trip, up and back by elevator	2	–	enclosed and glass	2628	258
Office4	round trip, up by elevator, back by stairs	4	–	enclosed	2047	199
Station1	one-way, down	1	7.29	glass	1346	131
Station2	round trip, down by elevator, back by ramp	1	–	glass	2987	291
Station3	round trip, up by elevator, back by stairs	1	–	glass	2179	212
Residential1	one-way, up	5	19.79	enclosed	1702	166
Residential2	one-way, up	15	49.98	enclosed	4104	401
Residential3	round trip, up and back by elevator	2	–	enclosed	5707	558
Campus1	one-way, up	2	13.65	enclosed	1420	138
Campus2	one-way, up	4	15.07	enclosed	2100	205
Campus3	one-way, up	5	15.97	enclosed	1456	142
Campus4	round trip, down by elevator, back by stairs	1	–	enclosed	2367	231
Campus5	round trip, up by elevator, back by stairs	1	–	enclosed	2666	260
Campus6	round trip, up and back by elevator	6	–	enclosed	3817	372

Simulated

Ten of the 16 sequences, none of them an elevator-free walk. Every tile is the ground-truth trajectory over the scene, with each ride drawn in orange and labelled with its rise. ● start, ■ end.

Sequence	Itinerary	Floors	H [m]	Elevator	Frames	Duration [s]
Office1-E	one-way, up	1	3.35	enclosed	1427	71
Office1-M	round trip, up and back by elevator	4	13.40	enclosed	3433	172
Office1-H	one-way, up	10	33.50	enclosed	2766	138
Office1-Walk	walk, no ride	–	–	–	2142	107
Office2-E	one-way, up	1	3.30	enclosed	1288	64
Office2-M	round trip, up and back by elevator	4	13.20	enclosed	3150	157
Office2-H	one-way, up	10	33.00	enclosed	2838	142
Office2-Walk	walk, no ride	–	–	–	2319	116
Mall-E	one-way, up	1	4.50	glass	1258	63
Mall-M	round trip, up and back by elevator	4	18.00	glass	5294	265
Mall-H	one-way, up	10	45.00	glass	3478	174
Mall-Walk	walk, no ride	–	–	–	2252	113
Factory-E	one-way, up	1	5.00	enclosed	1642	82
Factory-M	round trip, up and back by elevator	4	20.00	enclosed	4879	244
Factory-H	one-way, up	10	50.00	enclosed	3793	190
Factory-Walk	walk, no ride	–	–	–	1632	82

Download

Each sequence is two zips: <Seq>.zip (camera, IMU, reference trajectory, labels) and <Seq>_lidar.zip (LiDAR scans, about half the bytes). Camera+IMU only is 42 GB, everything 77 GB.

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 -n "$z"; done   # -> Elevator-VIGS/{real,sim}/<Seq>/

manifest.csv lists frames, scans, duration and checksums per sequence; sha256sum -c SHA256SUMS verifies a download. The code repository's scripts/prep_elevator.sh does the camera+IMU download for you.

Format

<arm>/<Seq>/               arm = real | sim
  images/<stamp>.jpg|png   real: <seconds>.jpg (1600x1296); sim: <nanoseconds>.png (640x480)
  imu.txt                  timestamp,gx,gy,gz,ax,ay,az   [rad/s, m/s^2]; real: seconds, one header line; sim: nanoseconds, no header
  calib.txt                fx fy cx cy   pinhole intrinsics, no distortion terms
  MT-Pose.txt              t x y z qw qx qy qz   body/IMU-frame pose, t in seconds on both arms
  lidar/<stamp>.npz        data (N,3) float32 xyz in the LiDAR frame, intensity (N,), timestamp (N,) uint64 ns per point; sim adds the scalars sim_time, sim_t0
  control_points.json      real only: scoring labels (see Evaluation)
  events.json              sim only: ride labels (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: the sensor clock (real) or the epoch 1754000000 s plus simulator time (sim). Sim file names and imu.txt carry it in nanoseconds, MT-Pose.txt in seconds.

Extrinsics. T_cb maps IMU/body-frame points into the camera frame, T_bl LiDAR-frame points into the body 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]]

IMU noise densities used in the paper (both arms): accelerometer 2.08e-3 m/s²/√Hz, random walk 4.13e-4; gyroscope 2.05e-4 rad/s/√Hz, random walk 1.11e-5.

Evaluation

Real-world, control_points.json. control_points maps each label to its still window [t0, t1] in seconds while the rig rests on the floor: B1 the departure floor, C1 the arrival floor (one-way), B2 the return to the departure floor (round trip). A method is scored on the median of its poses inside each window: the B1→C1 rise against H, and the B1→B2 height error against zero. Only heights along the method's own gravity direction are compared, so no trajectory alignment is needed. rides lists each ride as [t_start, t_end] from the IMU; the remaining keys are the provenance of these windows.

Simulated, events.json. legs lists every ride with shaft, from_floor, to_floor, rise and start/stop stamps; phases names the scripted segments (still, walk, boarding, ride, …); floor_height is the storey height. MT-Pose.txt also supports a full trajectory error.

The paper's scorers are in the code repository (scripts/elevator_eval_utils.py, eval_elevator_mono.py).

ROS bag

to_rosbag.py (needs numpy and rosbags, plus opencv-python for --images) builds a ROS1 bag from a sequence folder with /livox/lidar, /livox/imu and, with --images, /camera/image_raw. LiDAR-inertial methods disagree on the per-point time unit, hence the flag:

python to_rosbag.py Elevator-VIGS/real/Campus1                          # per-point time in seconds (e.g. Elevator-LIO)
python to_rosbag.py Elevator-VIGS/real/Campus1 --images --time-unit us  # microseconds + camera (e.g. FAST-LIVO2)

Privacy and license

The real-world sequences were recorded in public and residential buildings; faces are blurred in every released 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 sequences 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},
}