SReg: Effectively Registering Point Clouds of Different Scales

Point cloud registration (PCR) is to estimate a Sim(3) transformation, i.e., rotation, translation and scale, that aligns two point clouds. We study the challenging optimization problem of full-overlap registration for point clouds of different scales. Previous work commonly assumes the point clouds are of the same scale, and is susceptible to being trapped in a local minimum. We present SReg, a simple and training-free approach to effectively navigate the highly non-convex landscape of the problem. We propose three optimization strategies to substantially improve convergence effectiveness, namely decoupled Sim(3) transformation (to mitigate getting stuck in local minima), tangent space optimization (to overcome the discontinuities associated with large rotations), and coarse-to-fine bandwidth scheduling (to emphasize global geometric structure early on). Our extensive experiments demonstrate that SReg, compared with both SOTA learning-based and classical methods, achieves substantially lower Chamfer distance (by 69%), the highest registration recall (97%), stable registration under diverse scale and noise, and high efficiency in runtime and memory.