Collecting images and point cloud data of OoD objects is
challenging due to their rarity in real-world scenes and the
high cost of data collection and annotation. To address this,
we propose a Synthetic Anomaly Integration Pipeline that
generates synthetic anomalies under physical and environmental constraints, ensuring plausibility and challenge for
robust OoD detection model evaluation.
We applied the synthesis pipeline to the SemanticKITTI and SSCBench-KITTI-360 datasets, resulting in the creation of two new synthetic datasets: VAA-KITTI and VAA-KITTI-360. These synthesized datasets extend the original label sets by introducing anomaly labels, encompassing 26 distinct categories of anomalies such as animals, furniture, and garbage bags. .
- Refer to install.md to install the environment.
- Refer to dataset.md to prepare SemanticKITTI and KITTI360 dataset.
- Refer to run.md for training and evaluation.
- [2025/08] Code is now publicly available! The OOD (Out-of-Distribution) dataset will be released upon acceptance of the paper. Please stay tuned for updates.
- [2025/06] Init repository. The code and datasets will be made publicly available upon acceptance of the paper. Thank you for your interest in our work!
This project is based on the following open-source projects. We thank their authors for making the source code publically available.