Reweighted Discriminative Optimization for least-squares problems with point cloud registration

Optimization plays a pivotal role in computer graphics and vision. Learning-based optimization algorithms have emerged as powerful technique for solving problems with robustness accuracy because it learns gradients from data without calculating the Jacobian Hessian matrices. The key aspect of is least-squares method, which formulates general parametrized model unconstrained optimizations makes residual vector approach to zeros approximate solution. method may suffer undesirable local optima many applications, especially point cloud registration, where each element transformation vectors has different impact on registration. In this paper, Reweighted Discriminative (RDO) proposed. By assigning weights components parameter vector, RDO explores component asymmetrical contributions fitting results. are adjusted according characteristics mean square error results over space at per iteration. Theoretical analysis convergence provided, benefits demonstrated tasks 3D registrations multi-views stitching. experimental show that outperforms state-of-the-art registration methods terms perturbations achieves further improvement than non-weighting learning-based optimization.