C-GenReg: Training-free 3D point cloud registration using generative priors and VFMs

C-GenReg achieves zero-shot 3D point cloud registration by augmenting geometric matching with generative RGB views from a World Foundation Model, outperforming prior methods on indoor (3DMatch, ScanNet) and outdoor (Waymo) benchmarks without fine-tuning. The framework transfers point cloud matching to the image domain where VFMs excel, then lifts correspondences back to 3D. A 'Match-then-Fuse' scheme probabilistically combines geometric and RGB-derived correspondences, preserving modality biases. Limitations include reliance on pretrained models and potential failure when depth maps are noisy.

• #3d-registration
• #point-cloud
• #generative-models
• #vision-foundation-models
• #ben-gurion-university