Generalized-CVO
Fast and correspondence-free point cloud registration in RKHS with second-order Riemannian optimization
CVPR 2026 Highlight
Authors: Ray Zhang, Marcus Greiff, Thomas Lew, John Subosits
Links: Paper | Code | Demo video
Generalized-CVO estimates rigid SE(3) transforms between point clouds by maximizing a inner product with anisotropic kernels in a reproducing kernel Hilbert space. The method avoids explicit correspondences and combines geometric and appearance features with second-order Riemannian optimization on the SE(3) manifold.
Note: The original project website at sites.google.com/tri.global/gcvo is no longer maintained; this page now serves as the project page.
Overview of the correspondence-free formulation in RKHS, and the second order Riemannian solver.
Frame-to-frame registration examples from the GCVO repository: KITTI Seq 00 LiDAR odometry and ETH3D table_3 RGB-D tracking, both without local mapping or loop closure.
Demo of LiDAR localization using G-CVO with a prebuilt LiDAR map on an autonomous car.