BPNet: Bézier Primitive Segmentation on 3D Point Clouds

TL;DR

BPNet tackles the challenge of generalized parametric primitive segmentation for 3D point clouds by learning Bézier primitives instead of relying on a fixed repertoire of canonical shapes. It introduces a cascaded, end-to-end framework that jointly performs decomposition, fitting, embedding, and reconstruction, augmented by a soft voting regularizer and an auto-weight embedding module. The approach is validated on the ABC dataset and real scans, showing superior segmentation accuracy and faster inference than prior methods, while enabling full Bézier-based reconstruction with preserved sharp features. This work enables more flexible, edit-friendly representations of complex objects in 3D, with practical implications for modeling and visualization.

Abstract

This paper proposes BPNet, a novel end-to-end deep learning framework to learn Bézier primitive segmentation on 3D point clouds. The existing works treat different primitive types separately, thus limiting them to finite shape categories. To address this issue, we seek a generalized primitive segmentation on point clouds. Taking inspiration from Bézier decomposition on NURBS models, we transfer it to guide point cloud segmentation casting off primitive types. A joint optimization framework is proposed to learn Bézier primitive segmentation and geometric fitting simultaneously on a cascaded architecture. Specifically, we introduce a soft voting regularizer to improve primitive segmentation and propose an auto-weight embedding module to cluster point features, making the network more robust and generic. We also introduce a reconstruction module where we successfully process multiple CAD models with different primitives simultaneously. We conducted extensive experiments on the synthetic ABC dataset and real-scan datasets to validate and compare our approach with different baseline methods. Experiments show superior performance over previous work in terms of segmentation, with a substantially faster inference speed.

Figures (5)

• Figure 1: Overview of the proposed pipeline. The network takes a point cloud as input. It outputs pointwise features followed by four modules to predict Bézier geometry and topology: (1) The decomposition module decomposes point clouds into multiple patches. (2) The fitting module regresses $uv$ parameters for each point and control points for each Bézier patch. (3) The embedding module clusters pointwise features assigned to the same patch. (4) The reconstruction module re-evaluates the input point clouds from the above predictions.
• Figure 2: Comparisons on the ABC dataset, where GT-Ours denotes the ground truth for Bézier decomposition, and GT-Basic denotes the ground truth for HPNet, ParSeNet, and SPFN.
• Figure 3: Relaxed IOU loss with or without soft voting loss.
• Figure 4: Stress tests on real-scan data. The above two rows are CAD point clouds, and the last two are free-form point clouds.
• Figure 5: Reconstruction of the full Bézier model and visual comparison with ParSeNet.