Scalable and High-Quality Neural Implicit Representation for 3D Reconstruction

TL;DR

This work introduces Neural SDF-Graph, a divide-and-conquer framework for neural implicit 3D reconstruction that decomposes a scene into overlapping local SDFs, registers adjacent regions via a minimum spanning tree, and blends them with softmax-based weights to form a coherent global SDF. The approach achieves high-fidelity geometry while enabling scalable reconstruction of urban-scale scenes, demonstrated on datasets such as Lego, Jade, Sub-Campus, and Campus, with per-node texturing and editing capabilities. Key contributions include a graph-based representation of local SDFs, MST-driven registration to align local frames, and a seamless SDF blending strategy that mitigates seams at overlaps. The results show clear improvements in Chamfer distance and F-scores as the number of nodes increases, along with practical applications in texture generation and scene editing, highlighting the method's scalability and utility for large-scale 3D reconstruction.

Abstract

Various SDF-based neural implicit surface reconstruction methods have been proposed recently, and have demonstrated remarkable modeling capabilities. However, due to the global nature and limited representation ability of a single network, existing methods still suffer from many drawbacks, such as limited accuracy and scale of the reconstruction. In this paper, we propose a versatile, scalable and high-quality neural implicit representation to address these issues. We integrate a divide-and-conquer approach into the neural SDF-based reconstruction. Specifically, we model the object or scene as a fusion of multiple independent local neural SDFs with overlapping regions. The construction of our representation involves three key steps: (1) constructing the distribution and overlap relationship of the local radiance fields based on object structure or data distribution, (2) relative pose registration for adjacent local SDFs, and (3) SDF blending. Thanks to the independent representation of each local region, our approach can not only achieve high-fidelity surface reconstruction, but also enable scalable scene reconstruction. Extensive experimental results demonstrate the effectiveness and practicality of our proposed method.

Figures (15)

• Figure 1: Pipeline. At the bottom, we present the main construction processes of our method: Constructing the distribution and overlap relationship of the local radiance fields based on object structure or data distribution, Adjacent Nodes Registration, and SDF Blending. At the top, we take adjacent nodes $v_i$ and $v_j$ as an example to explain in detail how to use volume rendering for registration and then blending by the softmax-based weighting.
• Figure 2: Mesh Visualization for Optimization of Registration. The left image shows misalignment caused by the initial registration, while the right displays the optimized result.
• Figure 3: Blending. The top images depict simplified SDF contours within a vertical y-axis section. The left image shows a visible seam caused by directly taking the minimum value, while the right displays the smoothed result achieved through Softmax weighting.
• Figure 4: Lego Nodes Division. From left to right are the node divisions of $2$-nodes, $4$-nodes and $8$-nodes.
• Figure 5: Visualization of High-Quality Reconstruction Results for Lego. The right side is the comparison, demonstrating that our representation progressively reconstructs finer details.