BrepGaussian: CAD reconstruction from Multi-View Images with Gaussian Splatting

Jiaxing Yu; Dongyang Ren; Hangyu Xu; Zhouyuxiao Yang; Yuanqi Li; Jie Guo; Zhengkang Zhou; Yanwen Guo

cs.CVFeb 24, 2026

BrepGaussian: CAD reconstruction from Multi-View Images with Gaussian Splatting

Authors: Jiaxing Yu, Dongyang Ren, Hangyu Xu, Zhouyuxiao Yang, Yuanqi Li, Jie Guo, Zhengkang Zhou, Yanwen Guo

Organizations: State Key Laboratory of Novel Software Technology, Nanjing University · Nanjing Urban Construction Tunnel& Bridge Intelligent Management Co., Ltd.

Abstract

The boundary representation (B-Rep) models a 3D solid as its explicit boundaries: trimmed corners, edges, and faces. Recovering B-Rep representation from unstructured data is a challenging and valuable task of computer vision and graphics. Recent advances in deep learning have greatly improved the recovery of 3D shape geometry, but still depend on dense and clean point clouds and struggle to generalize to novel shapes. We propose B-Rep Gaussian Splatting (BrepGaussian), a novel framework that learns 3D parametric representations from 2D images. We employ a Gaussian Splatting renderer with learnable features, followed by a specific fitting strategy. To disentangle geometry reconstruction and feature learning, we introduce a two-stage learning framework that first captures geometry and edges and then refines patch features to achieve clean geometry and coherent instance representations. Extensive experiments demonstrate the superior performance of our approach to state-of-the-art methods.

Figures & tables

Figure 2: Overall pipeline of BrepGaussian . Given multi-view RGB images of a CAD object, we extract edge and patch views using existing edge detection and segmentation models. These views drive a two-stage Gaussian Splatting model that predicts edge and patch labels on the reconstructed point cloud. The fitted primitives are globally optimized to obtain the final B-Rep model.

Figure 3: Illustration of optimized 2D Gaussians. flat regions use fewer nearly spherical Gaussians, while edge regions require more elongated Gaussians.

Figure 4: Qualitative comparison on patch segmentation. Our BrepGaussian produces cleaner and more consistent patch segmentation.

Figure 5: Qualitative comparison on CAD reconstruction. Our BrepGaussian exhibit more accurate geometry.

Method	Input	Prec $\uparrow$	Rec $\uparrow$	F 1 $\uparrow$
ParSeNet [ 36 ]	GT Points	0.5108	0.2646	0.3486
PCER-Net [ 26 ]	GT Points	0.8760	0.9119	0.8936
SED-Net [ 25 ]	GT Points	0.9490	1.0000	0.9738
HPNet [ 40 ]	GT Points	1.0000	0.2142	0.3528
ParSeNet [ 36 ]	Densified	0.6229	0.2364	0.3427
PCER-Net [ 26 ]	Densified	0.5357	0.7924	0.6392

Table 1: Quantitative comparison of different methods for patch segmentation. The best results are shown in bold.

Method	Input	$\mathbf{D_{c}(10^{-2})\downarrow}$		$\mathbf{D_{h}(10^{-1})\downarrow}$
Method	Points	Surface	Curve	Surface	Curve
Point2CAD [ 29 ]	Our labels	3.38	5.42	2.413	3.858
Point2CAD [ 29 ]	PCER-Net	7.08	20.45	3.394	7.276
Split-and-Fit [ 27 ]	Densified	6.23	13.98	3.523	4.962
Ours	Our points	4.90	5.01	3.351	3.626

Table 3: Quantitative comparison of different methods for CAD reconstruction.

Pipeline	Prec $\uparrow$	Rec $\uparrow$	F1 $\uparrow$
w/o two-stage learning	0.8709	0.6805	0.7672
w/o triplet loss $\mathcal{L}_{tri}$	0.8395	0.9374	0.8857
Full Model	0.8903	0.9181	0.9040

Table 4: Ablation study on patch segmentation.

Data Preparation	$\mathbf{Prec\uparrow}$	$\mathbf{Rec\uparrow}$	$\mathbf{F_{1}\uparrow}$	$\mathbf{D_{c}(10^{-2})\downarrow}$
Manual Correction	0.8704	0.9191	0.8941	5.04
Fine-tuned SAM	0.8896	0.8363	0.8621	5.33
Vanilla SAM	0.8435	0.8090	0.8259	5.66

Table 6: Ablation study on mask quality.

Figure 6: From left to right, we present 2D images, point cloud and final B-Rep in 20, 30, 40 and 50 views.

Figure 7: Experiments on real-world scenes from ABO dataset [ 6 ] and smartphone photos.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 930 papers

UniBRep: Learning Unified Geometry and Topology for Image-conditioned B-Rep Generation
Oct 2, 2026Haiyang Ying, Allen Tu, Jiaye Wu +2Parametric CAD Modeling Single-View 3D Reconstruction
University of Maryland, College Park, Maryland, USA · Amazon, Bellevue, Washington, USA
GrapeSplat: Geometry-Grounded Reconstruction via Amalgamated Pose-Free Encoding for Feed-Forward 3D Gaussian Splatting
Sep 19, 2026Si-Yu Lu, Yung-Yao Chen, Yi Jan Chen +33D Gaussian Splatting 3D Reconstruction
National Taiwan University · National Taiwan University of Science and Technology · NTU AI Center of Research Excellence (NTU AI-CoRE) +1
StructSplat: Generalizable 3D Gaussian Splatting from Uncalibrated Sparse Views
Jun 26, 2026Jia-Chen Zhao, Beiqi Chen, Xinyang Chen +23D Gaussian Splatting 3D Reconstruction
Harbin Institute of Technology (Shenzhen) · Great Bay University · Guangzhou CloudButterfly Technology Co., Ltd.

Oct 2, 2026J/K move · Enter open · S save

UniBRep: Learning Unified Geometry and Topology for Image-conditioned B-Rep Generation

Haiyang Ying, Allen Tu, Jiaye Wu, Tom Goldstein+1

University of Maryland, College Park, Maryland, USA · Amazon, Bellevue, Washington, USA

Parametric CAD Modeling Single-View 3D Reconstruction

Open paper AskPDF arXiv

Generating a boundary representation (B-rep) conditioned on a single image requires faithful reconstruction of geometry, valid topology, and support for complex shapes. We present UniBRep, a geometry-first framework that adapts a pretrained image-to-3D model to generate a feature mesh as a unified intermediate representation. Its surface provides a geometric scaffold, while spatially aligned learned features encode face-separation cues for topology recovery. Dual decoder branches generate the geometry and face-separation features; a geometry- and feature-guided construction pipeline then fits parametric surfaces, recovers boundary curves and connectivity, and assembles an explicit B-rep using a CAD kernel. Recovering topology from mesh regions avoids predefined architectural face-count limits, allowing face count to scale with shape complexity. On the standard DeepCAD benchmark, UniBRep produces valid B-reps for 80.49% of inputs and reduces face Chamfer distance from 0.1096 to 0.0345 relative to CADDreamer. In a matched comparison, UniBRep also outperforms the HoLa public demo across all reported metrics. Further evaluations demonstrate scalability to high-complexity shapes beyond the standard 30-face range, generalization to objects outside the CAD training distribution, and qualitative transfer to real photographs.

Abstract

3D Reconstruction

Explore similar work

UniBRep: Learning Unified Geometry and Topology for Image-conditioned B-Rep Generation

GrapeSplat: Geometry-Grounded Reconstruction via Amalgamated Pose-Free Encoding for Feed-Forward 3D Gaussian Splatting

StructSplat: Generalizable 3D Gaussian Splatting from Uncalibrated Sparse Views