BeyondRPC: A Contrastive and Augmentation-Driven Framework for Robust Point Cloud Understanding

Beyond RPC is a framework combining 3D Point Cloud Adaptive Contrastive Learning with WOLFMix

🔍 Key Contributions

• ✅ Adaptive Contrastive Learning: Intra-modal and cross-modal contrastive losses with dynamic EWMA weighting.
• ✅ Pretraining with DGCNN: Using RGB or grayscale renderings as the secondary modality.
• ✅ Evaluation on Corrupted Data: Includes PointCloud-C evaluation on ModelNet40 and ShapeNetPart.

Paper Link

PDF

⚠️ Correction Note (June 2025):
In the published paper, the value reported as the mean Corruption Error (mCE) was mistakenly computed as the square root of the actual mCE.
The correct mCE is 0.59. This correction does not affect the ranking between models or the conclusions of the paper.

---

Acknowledgements

The zoo code is borrowed from PointCloud-C repository.