MergeNet: Explicit Mesh Reconstruction from Sparse Point Clouds via Edge Prediction

TL;DR

MergeNet reframes mesh reconstruction from sparse point clouds as an edge connectivity prediction problem to enable explicit mesh generation without costly iso-surface extraction. It introduces a three-stage pipeline—candidate edge generation, edge embedding through local canonical normalization, and edge-to-surface distance regression—followed by mesh assembly from filtered edges. The method achieves state-of-the-art performance among explicit approaches and competitive results with implicit methods, particularly under sparse data, while improving efficiency. Ablation studies validate the importance of canonical edge embedding, and the work points to future improvements in hole filling and broader applicability to point-cloud tasks.

Abstract

This paper introduces a novel method for reconstructing meshes from sparse point clouds by predicting edge connection. Existing implicit methods usually produce superior smooth and watertight meshes due to the isosurface extraction algorithms~(e.g., Marching Cubes). However, these methods become memory and computationally intensive with increasing resolution. Explicit methods are more efficient by directly forming the face from points. Nevertheless, the challenge of selecting appropriate faces from enormous candidates often leads to undesirable faces and holes. Moreover, the reconstruction performance of both approaches tends to degrade when the point cloud gets sparse. To this end, we propose MEsh Reconstruction via edGE~(MergeNet), which converts mesh reconstruction into local connectivity prediction problems. Specifically, MergeNet learns to extract the features of candidate edges and regress their distances to the underlying surface. Consequently, the predicted distance is utilized to filter out edges that lay on surfaces. Finally, the meshes are reconstructed by refining the triangulations formed by these edges. Extensive experiments on synthetic and real-scanned datasets demonstrate the superiority of MergeNet to SoTA explicit methods.

Figures (7)

• Figure 1: The illustration of the pipelines of MergeNet. (a): sparse input point cloud; (b): candidate edges (colors indicate different edges); (c): edges selected by MergeNet (thick lines); (d): refined mesh formed by filtered edges.
• Figure 2: Architecture of the proposed MergeNet which consists of three main steps: (a) Candidate edges generation module, (b) Edge embedding module and (c) Edge-to-surface distance regression loss.
• Figure 3: The illustration of local canonical normalization. (a) Example edge with endpoints $\boldsymbol{p}_s$ and $\boldsymbol{p}_t$. (b) Midpoint $\boldsymbol{c}_e$ of the edge as origin and the plane of $\mathbf{\boldsymbol{p}_s\boldsymbol{p}_t\boldsymbol{c}_g}$ as the $XY$ plane. (c) Canonical coordinate systems with left-hand rule.
• Figure 4: Comparison of reconstruction visualizations with baselines. Baselines are implicit reconstruction methods, including PSR, On-Surface, as well as explicit reconstruction methods such as BPA, IER and PointTriNet.
• Figure 5: Error map between the predicted edge-to-face distance and GT. Colder color indicates smaller error.