Non-rigid point cloud registration with dynamic computing neural deformation pyramid
Jiankun Huo, Jiahui Huang, Yinling Qian, Qiong Wang · 2025
Non-rigid point cloud registration plays a crucial role in computer vision, robotic perception, and medical image analysis. However, existing methods often employ fixed network architectures when processing point clouds with varying deformation complexities, disregarding the heterogeneity of point cloud data. This limitation constrains both registration accuracy and generalization capability. To address this issue, we propose DCNDP (Dynamic Computing Neural Deformation Pyramid), a non-rigid point cloud registration method that dynamically adjusts the depth of deformation modeling based on point cloud complexity. DCNDP computes the complexity of the point cloud and adaptively selects the hierarchical levels of the NDPLayer deformation pyramid to accommodate different deformation requirements. Experimental results on the 4DMatch dataset demonstrate that DCNDP outperforms existing methods in terms of registration accuracy, exhibiting superior stability and robustness, particularly in scenarios involving large and complex deformations.