Event
CVSD-Reg Distills DINOv2 Semantics for Zero-Shot Cross-Sensor LiDAR Registration
1 reports · 1 sources
Eunsoo Im et al. propose the CVSD-Reg framework, which distills frozen DINOv2 visual semantic priors into a Point Transformer V3 student network, achieving strict success rates (SR@0.5m/1°) of 97.7%, 99.0%, and 99.3% on KITTI, nuScenes, and HeLiPR, including 97.3% on sparse 16-beam Velodyne, leading state-of-the-art geometric registration methods by up to 44.0 percentage points. At inference, only LiDAR point clouds and a single checkpoint are used to perform descriptor extraction, correspondence matching, and pose estimation, with no camera input, sensor-specific adaptation, or post-hoc ICP refinement required; training proceeds in two stages, first cross-modal distillation to obtain robust descriptors, then fine-tuning on the registration task.
Report timeline
- CVSD-Reg Distills DINOv2 Semantics for Zero-Shot Cross-Sensor LiDAR Registration
CVSD-Reg is a robust global LiDAR registration framework that distills frozen DINOv2 visual semantic priors into a Point Transformer V3 student network, achieving zero-shot cross-sensor generalization with no camera needed at inference time.
3D视觉工坊 · 微信公众号 ·