Reproduce “Neural-IMLS: Self-supervised Implicit Moving Least-Squares Network for Surface Reconstruction” on OpenResearch

About this paper

Zixiong Wang, Pengfei Wang, Pengshuai Wang, Qiujie Dong, Junjie Gao, Shuangmin Chen, Shiqing Xin, Changhe Tu and 1 more

Neural-IMLS reconstructs high-fidelity 3D surfaces from noisy, unoriented point clouds by introducing a self-supervised mutual regularization between a Multilayer Perceptron (MLP) and an Implicit Moving Least-Squares (IMLS) function. This method, developed by researchers from Shandong University, Peking University, Qingdao University of Science and Technology, and Texas A&M University, consistently outperforms prior self-supervised approaches on various benchmarks, demonstrating improved robustness and detail preservation.

Key insights

• A self-supervised mutual regularization mechanism between a Multilayer Perceptron (MLP) and an Implicit Moving Least-Squares (IMLS) function enables robust surface reconstruction without ground-truth data or pre-computed normals.
• The MLP's gradients provide normal estimates for IMLS, while IMLS's output serves as a dynamic, noise-resistant target for the MLP, leading to iterative refinement.
• A novel spatial coherence loss, based on normalized gradient directions, is introduced within the IMLS weighting scheme to preserve sharp features and reduce artifacts.

Summary by alphaXiv.