3D Common Corruptions

• Publisher: EPFL VILAB
• Status: active
• Version: N/A
• Release Date: 2022-09-01
• Date Added: 2026-01-07
• Source URL: https://github.com/EPFL-VILAB/3DCommonCorruptions

---

Summary

3D Common Corruptions is a research benchmark and dataset for evaluating the robustness of 3D vision models under systematic corruption and distribution shift. It extends the concept of “common corruptions” from 2D vision to 3D modalities such as point clouds and meshes, enabling controlled stress-testing of perception systems. The tool is primarily used in robustness evaluation, model validation, and research on reliability of 3D perception in safety- and security-relevant contexts.

---

Key Takeaways

• Provides standardized corruption types that can be applied to most datasets, even those that do not contain 3D information.
• Enables robustness and generalization testing of 3D vision models beyond clean benchmarks
• Enables realistic augmentation for training data to improve model generalization
• Supports research into failure modes and distribution shift for perception systems

---

Related Code

NA

---

Additional Sources

• EPFL VILAB Publications

---

License

MIT