Shellbe: 3D-LiDAR centered multi-sensor dataset in a structurally degraded environment.

📚 Table of Contents

• Competition Links

• Dataset Description

• Dataset Descriptions

  • Public Training Datasets
  • Validation Datasets
  • Testing Datasets

• Data Structure and File Organization

  • Reference Contents

• Downloads

• Examples and Teasers

  • Video Teasers

Datasets and Technical Details

Competition Links

• Official Competition Platform: https://comrob-ds.fel.cvut.cz:555/competitions/18/
• Evaluation Framework (slam-bench): https://github.com/comrob/slam-bench/tree/main

---

Dataset Description

Visualization of dataset collection areas.

The datasets consist of multiple recorded loops and calibration sequences intended for evaluation and training of sensor-based localization and mapping. They were recorded outdoors in challenging localization conditions. The datasets include LiDAR, GNSS, multiple cameras (RGB, thermal), and IMUs. NIR and NDVI are provided by the Agiception camera.

---

Dataset Descriptions

Public Training Datasets

• shellby-0225-train-loop1 (451 m)   Loop in an open field used for training, moving further from trees.

• shellby-0225-train-lab   Short indoor recording from CTU Computational Robotics Lab for initial testing.   Uses Total Station instead of GNSS.

Validation Datasets

• shellby-0225-validation-loop1 (313 m)   Small loop primarily for testing submissions. The forest remains in LiDAR range.

Testing Datasets

• shellby-0225-test-loop1 (1892 m)   Long loop with both field and forest. Includes a 30-second LiDAR outage due to power loss.

• shellby-0225-test-loop2 (667 m)   Similar to the training loop but with less smooth trajectory.   Evaluated using Total Station data. Basler camera slightly overexposed.

---

Data Structure and File Organization

The dataset is organized to be compatible with the slam-bench evaluation framework.

<dataset\_name>/
├── calibration/
│   ├── extrinsics/
│   │   ├── extrinsics.txt
│   │   ├── static_tf.launch
│   │   └── static_tf.urdf
│   └── instrinsics/
│       ├── basler.yaml
│       └── ...
├── reference/
│   ├── reference.txt
│   └── ...
├── sensors/
│   └── <all_bagfiles>.bag
└── tracks/
   ├── all.txt
   └── passive.txt

• sensors/<all_bagfiles>.bag → Raw sensory data in ROS bag files sequence.
• calibration/ → Contains intrinsic and extrinsic calibration parameters.
• reference/ → Folder containing the ground-truth reference trajectory.
• tracks/ → Contains files defining subsets of ROS topics for specific evaluation scenarios (e.g., passive.txt might only list topics from non-interfering sensors). The default track is all, which plays all sensors.

Reference Contents

Each dataset contains a reference/ subdirectory with:

• reference.txt: The ground-truth trajectory in TUM format.

---

Downloads

Training datasets are provided on GoogleDrive.

---

Examples and Teasers

Example environments where data was collected.

Video Teasers

Train Dataset

Test Datasets