Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving

Alexey Nekrasov¹, Malcolm Burdorf¹, Stewart Worrall², Bastian Leibe¹, Julie Stephany Berrio Perez²

¹RWTH Aachen University ²The University of Sydney

3D Anomaly Segmentation Dataset

[Project Webpage] [Paper]

---

🗞️ News

• 2025-06-14: Images Released
• 2025-05-29: Training Code and Checkpoints Released
• 2025-03-25: Data and Evaluation Code Released
• 2025-02-26: STU Accepted at CVPR 2025

---

💽 Data

• STU dataset is available at STU.
• PANOPTIC-CUDAL dataset is available at Panoptic-CUDAL.

To verify that downloaded files are correct, you can verify the SHA256 hash of the files.

sha256sum -c file_sha256sum.chk

Overall the data follows the SemanticKITTI format.

|── 125/
|   ├── poses.txt
|   ├── calib.txt
|   ├── labels/
|   │     ├ 000000.label
|   │     └ 000001.label
|   .
|   |
|   └── velodyne/
|         ├ 000000.bin
|         └ 000001.bin
.
.
└── 134/

Predictions are simple .txt files with confidence per point.

🏁 Evaluation

Simple evaluation for point-level anomaly segmentation:

python compute_point_level_ood.py --data-dir stu_dataset/val --pred-dir ./prediction

Simple evaluation for point-level anomaly segmentation:

python compute_object_level_ood.py --data-dir stu_dataset/val --instance-dir ./instance_prediction

🏎️ Training and Inference

Please check Mask4Former3D folder in the repository

📒 TODO

• Open Public Test Submission
• Release anonymized images
• Release training code and checkpoints
• Release code for points projection to images
• Release the data
• Release evaluation code

As of June 14, 2025, the public test submission is delayed due to the size of the test set and the computing requirements for evaluation. I had initially planned to host the test set on Codalab/Codabench; however, a single submission for point-level evaluation would require a 3 GB upload, which is too large and takes too long to evaluate. I will try to find a workaround for this limitation. In the meantime, please feel free to evaluate on validation set, and if you have a method you would like to evaluate on test, email me and we can figure it out.

🙏 Acknowledgement

Many thanks to reviewers of our paper submission. You helped us improve the project a lot.

BibTeX

@inproceedings{nekrasov2025stu,
  title = {{Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving}},
  author = {Nekrasov, Alexey and Burdorf, Malcolm and Worrall, Stewart and Leibe, Bastian and Julie Stephany Berrio Perez},
  booktitle = {{"Conference on Computer Vision and Pattern Recognition (CVPR)"}},
  year = {2025}
}