This paper presents an efficient variant of the ICP (Iterative Closest Point) algorithm for aligning three-dimensional models. The ICP algorithm is widely used for geometric alignment when an initial estimate of the relative pose is known. The authors classify and evaluate various ICP variants, focusing on their impact on convergence speed. They introduce a new variant based on uniform sampling of the space of normals, which improves convergence for nearly-flat meshes with small features. The paper also proposes a combination of ICP variants optimized for high speed, demonstrating an implementation that can align two range images in a few tens of milliseconds, assuming a good initial guess. This capability has potential applications in real-time 3D model acquisition and model-based tracking. The authors compare different ICP variants across six stages: point selection, point matching, weighting, pair rejection, error metric, and minimization. They find that the point-to-plane error metric performs better than the point-to-point metric, especially for scenes with sparse features. The paper also discusses the effects of various weighting strategies, pair rejection methods, and error metrics on convergence speed. The authors conclude that a high-speed ICP algorithm can be constructed by combining several variants, resulting in a method that aligns two meshes in a few tens of milliseconds. The paper also includes an appendix discussing scanner noise and weighting.This paper presents an efficient variant of the ICP (Iterative Closest Point) algorithm for aligning three-dimensional models. The ICP algorithm is widely used for geometric alignment when an initial estimate of the relative pose is known. The authors classify and evaluate various ICP variants, focusing on their impact on convergence speed. They introduce a new variant based on uniform sampling of the space of normals, which improves convergence for nearly-flat meshes with small features. The paper also proposes a combination of ICP variants optimized for high speed, demonstrating an implementation that can align two range images in a few tens of milliseconds, assuming a good initial guess. This capability has potential applications in real-time 3D model acquisition and model-based tracking. The authors compare different ICP variants across six stages: point selection, point matching, weighting, pair rejection, error metric, and minimization. They find that the point-to-plane error metric performs better than the point-to-point metric, especially for scenes with sparse features. The paper also discusses the effects of various weighting strategies, pair rejection methods, and error metrics on convergence speed. The authors conclude that a high-speed ICP algorithm can be constructed by combining several variants, resulting in a method that aligns two meshes in a few tens of milliseconds. The paper also includes an appendix discussing scanner noise and weighting.