Point2Quad: Generating Quad Meshes from Point Clouds via Face Prediction

TL;DR

This work tackles quad mesh generation directly from point clouds by reframing quadization as a local face prediction task. Point2Quad combines a KPConv-based point encoder with a face encoder that captures topological cues, generates candidate quadrilaterals via k-NN, and uses a compound loss to train a 2-class face classifier, followed by quad-specific post-processing to enforce watertight, manifold quad meshes. It introduces a Point&Quad dataset and demonstrates superior quantitative and qualitative performance against baselines under noise and varying point resolutions, while conducting extensive ablations to validate each component. The approach advances practical quad mesh reconstruction with robust surface fidelity and topological quality, though it acknowledges limits in extreme noise or sparse inputs and suggests directions for improved generalization and manifold guarantees.

Abstract

Quad meshes are essential in geometric modeling and computational mechanics. Although learning-based methods for triangle mesh demonstrate considerable advancements, quad mesh generation remains less explored due to the challenge of ensuring coplanarity, convexity, and quad-only meshes. In this paper, we present Point2Quad, the first learning-based method for quad-only mesh generation from point clouds. The key idea is learning to identify quad mesh with fused pointwise and facewise features. Specifically, Point2Quad begins with a k-NN-based candidate generation considering the coplanarity and squareness. Then, two encoders are followed to extract geometric and topological features that address the challenge of quad-related constraints, especially by combining in-depth quadrilaterals-specific characteristics. Subsequently, the extracted features are fused to train the classifier with a designed compound loss. The final results are derived after the refinement by a quad-specific post-processing. Extensive experiments on both clear and noise data demonstrate the effectiveness and superiority of Point2Quad, compared to baseline methods under comprehensive metrics.

Figures (10)

• Figure 1: Full pipeline of our Point2Quad: given a point cloud as input, we first propose a set of candidate quadrilateral. During training (dashed arrow), the network is trained to classify the candidates with the supervision of face labels. During inference, the predicted label is used to filter out quadrilaterals, which are then merged into the output mesh with a post-process.
• Figure 2: Illustration of notions in quadrilaterals.
• Figure 3: post-process: ① remove non-manifold parts with scores in Eq. \ref{['eq:FaceScore']}; ② fill holes automatically. Non-manifold and boundary edges are highlighted in blue and red.
• Figure 4: Scoring example: The blue and red edges are non-manifold and boundary edges respectively.
• Figure 5: Fillhole strategy in greedy algorithm