Automated 3D Dataset Generation for Arbitrary Objects
6D pose estimation is a cornerstone of various autonomous systems, including mobile robots and self-driving cars, because it enables fundamental capabilities such as collision-free navigation and object grasping. However, training performant pose estimation models requires large and precisely annotated datasets. Due to the complex, expensive, and time-consuming capturing and annotation process, only few 3D datasets are available, which are also often limited in class diversity. To provide practitioners with a tool capable of generating 3D datasets for arbitrary objects, we propose an end-to-end pipeline that automates all aspects of dataset generation. By leveraging the implicit modeling capabilities of Radiance Fields, our pipeline constructs high-quality 3D models for arbitrarily complex objects. These 3D models serve as input for a synthetic dataset generator. A comprehensive evaluation across multiple objects demonstrates that our pipeline is fast, easy to use, and highly automated. Furthermore, we show that pose estimation networks trained on the generated datasets achieve strong performance in typical application scenarios
Automated 3D Dataset Generation for Arbitrary Objects
6D pose estimation is a cornerstone of various autonomous systems, including mobile robots and self-driving cars, because it enables fundamental capabilities such as collision-free navigation and object grasping. However, training performant pose estimation models requires large and precisely annotated datasets. Due to the complex, expensive, and time-consuming capturing and annotation process, only few 3D datasets are available, which are also often limited in class diversity. To provide practitioners with a tool capable of generating 3D datasets for arbitrary objects, we propose an end-to-end pipeline that automates all aspects of dataset generation. By leveraging the implicit modeling capabilities of Radiance Fields, our pipeline constructs high-quality 3D models for arbitrarily complex objects. These 3D models serve as input for a synthetic dataset generator. A comprehensive evaluation across multiple objects demonstrates that our pipeline is fast, easy to use, and highly automated. Furthermore, we show that pose estimation networks trained on the generated datasets achieve strong performance in typical application scenarios
Citing
```bibtex @article{schulz2025automated, title={Automated 3D Dataset Generation for Arbitrary Objects}, author={Schulz, Paul and Hempel, Thorsten and Al-Hamadi, Ayoub}, journal={IEEE Access}, year={2025}, publisher={IEEE} }