High-Fidelity Surface Splatting-Based 3D Reconstruction from Multi-View Images

2026-05-08

AI Summary

This work addresses the key challenge in 3D surface reconstruction from sparse multi-view images: recovering high-frequency geometric details while jointly optimizing geometry and appearance. To this end, the authors propose an implicit moving least squares (IMLS) framework based on a compactly supported polynomial kernel, replacing conventional exponential kernels to enhance the flexibility and controllability of geometric representation. They further introduce stochastic regularization and Laplacian filtering to improve optimization stability and preserve fine-scale geometric details. The resulting method enables end-to-end differentiable surface reconstruction and rendering, achieving state-of-the-art performance across multiple benchmarks with significant gains in both geometric accuracy and visual sharpness.

Abstract

Multi-view mesh reconstruction remains a core challenge in computer graphics and vision, especially for recovering high-frequency geometry from sparse observations. Recent methods such as 3D Gaussian Splatting (3DGS) and Neural Radiance Fields (NeRF) rely on post-processing for mesh extraction, thereby limiting joint optimization of geometry and appearance. Implicit Moving Least Squares (IMLS) instead enables direct conversion of point clouds into signed distance and texture fields, supporting end-to-end reconstruction and rendering. However, existing IMLS formulations use exponential kernels that struggle with high-frequency detail. We introduce a compact polynomial kernel with local support and greater flexibility, allowing better control over frequency content and improved geometric fidelity. To further enhance fine details, we incorporate stochastic regularization with Laplacian filtering. Together, these improve the preservation of high-frequency structure while maintaining stable optimization. Experiments show state-of-the-art performance in both surface reconstruction and rendering, yielding more accurate geometry and sharper visuals from multi-view data.

Problem

• 3D reconstruction
• multi-view images
• high-frequency geometry
• surface splatting
• geometric fidelity

Innovation

• Implicit Moving Least Squares
• polynomial kernel
• high-frequency geometry
• stochastic regularization
• surface reconstruction