Search papers, labs, and topics across Lattice.
This paper introduces CVSD-Reg, a novel global LiDAR registration framework that leverages visual semantic priors from a vision foundation model to enhance robustness against variations in point cloud characteristics. By employing a two-stage approach that includes contrastive distillation and spherical-manifold alignment, the method preserves the geometry of the embedding space while ensuring viewpoint invariance. The results demonstrate that CVSD-Reg achieves exceptional registration success rates across multiple datasets, significantly outperforming traditional geometric methods without the need for camera inputs or additional refinement steps.
CVSD-Reg achieves a remarkable 97.7% success rate on challenging LiDAR datasets, outperforming existing methods by up to 44% without relying on camera data.
Learning-based global point cloud registration has achieved remarkable progress, yet its reliance on geometric representations makes existing methods sensitive to variations in point density, scan pattern, viewpoint, and sensor characteristics. We propose CVSD-Reg, a robust global LiDAR registration framework that distills visual semantic priors from a vision foundation model into LiDAR representations. In Stage 1, a Point Transformer V3 student learns from a frozen DINOv2 teacher through contrastive distillation and spherical-manifold alignment, which preserves the hyperspherical geometry of the teacher embedding space. Self-supervised InfoNCE consistency and soft $\mathrm{SE}(3)$ invariance further encourage viewpoint-robust descriptors. In Stage 2, the distilled representation is adapted to registration through correspondence learning, density-aware point-dropout augmentation, and end-to-end pose optimization. With a single checkpoint, CVSD-Reg generalizes to both single-sensor and zero-shot cross-sensor scenarios without sensor-specific adaptation and remains entirely camera-free at inference. On KITTI, nuScenes, and HeLiPR, CVSD-Reg achieves strict success rate (SR@0.5\,m/$1^\circ$) of 97.7$\%$, 99.0$\%$, and 99.3$\%$, respectively, including 97.3$\%$ on sparse 16-beam Velodyne scans. It outperforms state-of-the-art geometric registration methods by up to 44.0 percentage points without requiring camera inputs or post-hoc ICP refinement.
