Scalable and Differentiable Point-Cloud Registration Using Maximum Mean Discrepancy
Provides a scalable and differentiable registration layer for end-to-end trainable models, addressing the need for robust registration in challenging conditions.
MMD-Reg is a differentiable, correspondence-free point-cloud registration method with linear complexity, achieving competitive accuracy and scalability against both learning-based and non-learning-based methods, especially under poor initial alignment and partial overlap.
We present MMD-Reg, a novel correspondence-free approach to point-cloud registration that is differentiable and has linear computational complexity in the number of points. We model registration as a nonlinear least-squares problem based on the Maximum Mean Discrepancy, approximated using random Fourier features. The resulting objective can be solved efficiently with standard methods such as Levenberg-Marquardt, and the solution is differentiable via the implicit function theorem. This allows MMD-Reg to be used as a differentiable optimization layer within end-to-end trainable models, supporting registration under challenging conditions such as poor initial alignment and partial overlap. We demonstrate this Neural MMD-Reg formulation by integrating the layer with a set transformer, training the resulting model in supervised and unsupervised settings, and comparing its performance against recent learning-based methods. We also evaluate standalone MMD-Reg, comparing its accuracy and scalability against widely used non-learning-based registration methods.