KevinWong216/MirrorSentinel-Elevator

MirrorSentinel Reflective Elevator Dataset

MirrorSentinel-Elevator is a synchronized RGB-LiDAR-IMU ROS 2 dataset for studying severe mirror and glass interference in compact elevator environments. It contains eight short traversals, E01-E08, collected across seven measured physical elevator footprints. E02 and E03 repeat the same physical cabin and are grouped as one footprint for inferential statistics.

The release contains the exact raw bags used by the accompanying paper. It does not contain generated maps, monocular-depth priors, reflection masks, learned-model weights, map-cleaning outputs, or a cabin ROI. The geometric annotations are evaluation-only and are never read by the mapping or cleaning pipeline.

• Repository: KevinWong216/MirrorSentinel-Elevator
• Version: 1.0.2
• Release date: 2026-07-27
• Authors: Yugong Wang, Shengzhe Gan, Qi Shen, and Sirun Pan
• Code: KvnWong216/MirrorSentinel
• Code revision used for this release: 27c4563605eeabc7926b6df1e80529c67d9adf7e
• License: Creative Commons Attribution 4.0 International

Dataset Summary

Property	Value
Sequences	8 synchronized ROS 2 bags
Physical footprints	7; E02 and E03 share one cabin
Total database size	30.61 GiB (32870649856 bytes)
Total duration	265.227 s
Modalities	Rectified left RGB, 3D LiDAR, IMU, camera calibration, TF
Storage	ROS 2 SQLite3 (.db3), metadata version 8
Audio	None
Splits	Evaluation collection; no prescribed train/validation/test split
Ground reference	Manually measured rectangular cabin geometry

Sequence Inventory

Sequence	Duration [s]	DB3 [GiB]	RGB	LiDAR	IMU	SHA-256 prefix
E01	39.266	4.550	1159	393	3927	3a24e137af3065d4
E02	35.254	4.081	1041	352	3526	2a245ca2fe3260a0
E03	31.002	3.601	920	310	3100	584dfceaaf3b5529
E04	26.777	3.091	787	267	2678	5bac233c9ff1e6eb
E05	22.537	2.608	665	225	2254	13d44f4f31e672b5
E06	40.026	4.558	1142	400	4003	1051dcc1d5bfe2ad
E07	23.167	2.686	684	232	2317	f3efb53a5f36ca46
E08	47.198	5.438	1378	472	4720	741f550197df0ce5

Full database hashes, exact nanosecond durations, topic types, and validation records are stored in manifest.json. SHA256SUMS covers every distributed file except itself.

Directory Layout

MirrorSentinel-Elevator/
  .gitattributes
  README.md
  LICENSE.md
  CITATION.cff
  RELEASE.json
  SHA256SUMS
  manifest.json
  data/
    elevator_01/
      elevator_01.db3
      metadata.yaml
    ...
    elevator_08/
      elevator_08.db3
      metadata.yaml
  annotations/
    elevator_01/room_gt_annotation.yaml
    ...
    elevator_08/room_gt_annotation.yaml

Each sequence contains the following topics:

Topic	ROS 2 type	Role
/ouster/points	sensor_msgs/msg/PointCloud2	3D LiDAR returns
/ouster/imu	sensor_msgs/msg/Imu	Inertial measurements
/zed2/zed_node/left/image_rect_color	sensor_msgs/msg/Image	Rectified left RGB
/zed2/zed_node/left/camera_info	sensor_msgs/msg/CameraInfo	Camera calibration
/tf	tf2_msgs/msg/TFMessage	Dynamic transforms
/tf_static	tf2_msgs/msg/TFMessage	Static sensor transforms

Camera calibration and sensor transforms required to interpret the streams are therefore embedded in every bag.

Download and Playback

Install a current Hugging Face client:

python3 -m pip install -U "huggingface_hub>=0.32"
hf download KevinWong216/MirrorSentinel-Elevator \
  --repo-type dataset \
  --revision 1.0.2 \
  --local-dir MirrorSentinel-Elevator

Verify all released bytes:

cd MirrorSentinel-Elevator
sha256sum --check SHA256SUMS

Inspect or replay one sequence with ROS 2 Humble:

ros2 bag info data/elevator_01
ros2 bag play data/elevator_01 --clock

The source-code repository includes the frozen mapper integration, visual prior adapters, map cleaner, evaluation implementation, and paper-facing experiment records. Follow its full-data reproduction instructions after placing this dataset at the documented data root.

Annotations and Evaluation Protocol

Each room_gt_annotation.yaml records one manually selected cabin corner, in-plane orientation, measured length and width, and a declared vertical interval. The file expands this compact rectangle into wall, floor, and ceiling geometry used by the released evaluator.

These annotations are not laser-scan ground truth and are not trajectory ground truth. They support cabin-boundary map-quality evaluation only. The informational bag: field retains a pre-release data-root alias for byte-level provenance; evaluation does not open that field.

The collection does not prescribe machine-learning splits. All E01-E08 sequences form the paper's target-domain evaluation collection. Researchers creating train/test partitions should report them explicitly and keep E02 and E03 together when independence at the physical-footprint level matters.

Intended Use

Appropriate uses include:

• reflective-surface robustness evaluation for LiDAR, LIO, and LIVO systems;
• multipath and behind-surface endpoint analysis;
• visibility-based map cleaning;
• RGB-LiDAR-IMU synchronization and fusion research;
• reproducibility studies of the accompanying MirrorSentinel paper.

The dataset was not designed or validated for face recognition, identity inference, surveillance, access-control analysis, or profiling of people or facilities. Incidental visual appearances do not constitute labels for those tasks.

Limitations

• The collection is deliberately focused: eight short elevator traversals over seven physical footprints.
• Cabin geometry is manually measured rather than obtained from a survey-grade 3D scanner.
• No globally referenced trajectory ground truth is distributed.
• Reflective layouts, materials, lighting, and sensor placement do not cover every elevator or glass environment.
• ROS 2 bags are domain-native research artifacts and are not rendered by the Hugging Face tabular dataset viewer.

Claims about new methods should therefore remain specific to the evaluated conditions or be supported by additional independent data.

Release, Privacy, and Rights

The maintainers reviewed the recordings and cleared this release for public distribution. The bags contain no audio. RGB frames can show the sensor operator, reflections, and institutional indoor spaces. CC BY 4.0 does not waive privacy, publicity, personality, trademark, or other third-party rights that may apply; downstream users remain responsible for their use.

The dataset is distributed under CC BY 4.0. Attribution is required. The license does not imply endorsement by the authors or their institution, and it does not grant rights to use names, logos, or trademarks for endorsement.

Citation

When using the released bags or annotations, cite the versioned dataset record:

@misc{wang2026mirrorsentinel_dataset,
  author       = {Yugong Wang and Shengzhe Gan and Qi Shen and Sirun Pan},
  title        = {MirrorSentinel Reflective Elevator RGB-LiDAR-IMU Dataset},
  year         = {2026},
  version      = {1.0.2},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/KevinWong216/MirrorSentinel-Elevator}},
  note         = {Eight ROS 2 elevator traversals over seven physical footprints}
}

For questions and reproducibility reports, use the MirrorSentinel issue tracker.