Point Cloud to Wireframe

Knowledge Base — 32 Papers, 6 Research Focus Areas, 1 Cross-Cutting Bottleneck Thread

Key Themes
Implicit Neural Representation Point Cloud Encoding Wireframe Reconstruction B-Rep / CAD Generation Self-Supervised 3D Learning 3D Diffusion Models Set-Structured Data Graph Neural Networks for 3D
Dual-Thread Reading

Primary thread: the timeline remains grouped by research focus area. Secondary thread: each category card is also tagged by the main bottleneck it attacks on the path from raw points to editable structure.

Feature acquisition Local geometric representation Discrete topology Validity constraints Generative priors Set / graph mechanisms Structure-aware evaluation
2019–2025

A. Point Cloud → Wireframe / Structured Geometry (Core Pipeline)

Directly relevant to the "point cloud to wire frame" problem. These papers tackle vertex detection, edge/topology reconstruction, B-Rep generation from point clouds, and end-to-end wireframe parsing — the core pipeline from raw 3D data to structured geometric output. In the second thread, this is the point where the literature stops treating shape as only continuous geometry and starts confronting discrete topology and validity explicitly.

Second thread Discrete topology Validity constraints Editable structure
PC2WF: 3D Wireframe Reconstruction from Raw Point Clouds
arXiv:2103.02766 (2021)
3D wireframe reconstruction from raw point clouds — vertex detection, localization, and edge detection pipeline.
3D Wireframe Reconstruction Vertex Detection Edge Detection Corner Localization Point Cloud to CAD
ComplexGen: CAD B-Rep Chain Complex Generation
arXiv:2205.14573 (2022)
CAD B-Rep chain complex generation from point clouds — joint vertex/edge/face + topology reconstruction.
B-Rep Chain Complex Joint Topology Reconstruction Vertex/Edge/Face Detection CAD Reverse Engineering Point Cloud to CAD
CLR-Wire: Continuous Latent Representation for 3D Curve Wireframes
arXiv:2504.19174 (2025)
3D curve wireframe generation — continuous latent representation for geometry+topology, conditions on point clouds.
3D Curve Wireframe Continuous Latent Representation Geometry-Topology Unification Neural Parametric Curve Conditional Wireframe Generation
L-CNN: End-to-End Wireframe Parsing
arXiv:1905.03246 (2019)
End-to-end 2D wireframe parsing (from images) — junction + line detection with LoI pooling.
Wireframe Parsing Junction Detection Line-of-Interest Pooling Structural AP End-to-End Vectorized Output
BrepGen: B-Rep Generative Diffusion Model
arXiv:2401.15563 (2024)
B-Rep generative diffusion model — directly generates CAD B-Rep with structured latent geometry.
B-Rep Diffusion Model Structured Latent Geometry CAD Generation Face/Edge/Loop Topology Boundary Representation
2017–2022

B. Point Cloud Feature Learning & Encoding

Learning compact vector representations and features from raw point clouds. This line of work spans autoencoders (FoldingNet, TearingNet), Transformers (Point Transformer, PCT, Point-BERT), and self-supervised masked autoencoders (Point-MAE, Point-M2AE). In the second thread, this block attacks the upstream bottleneck: how much local geometric signal can be extracted from unordered points before any structured prediction is attempted.

Second thread Feature acquisition Local geometric bias Self-supervision
Achlioptas et al.: Learning Representations and Generative Models for 3D Point Clouds
arXiv:1707.02392 (2017)
Point cloud autoencoders + generative models (GANs in latent space, GMMs) — pioneering raw point cloud learning.
Point Cloud Autoencoder Latent Space GAN Chamfer Distance 3D Generative Model Shape Interpolation
FoldingNet: Point Cloud Auto-encoder via Deep Grid Deformation
arXiv:1712.07262 (2017)
Point cloud AE via 2D grid folding decoder; unsupervised feature learning through reconstruction.
2D Grid Folding Unsupervised Feature Learning Point Cloud Reconstruction Chamfer Distance Grid Deformation
TearingNet: Point Cloud Autoencoder with Topology Adaptation
arXiv:2006.10187 (2020)
Topology-friendly point cloud AE — tears 2D grid to match holes and multi-object topology in reconstruction.
Topology-Aware Autoencoder 2D Grid Tearing Point Cloud Reconstruction Manifold Structure Multi-Object Topology
Point Transformer: Self-Attention Networks for 3D Point Cloud Processing
arXiv:2012.09164 (2020)
Self-attention layers for 3D point cloud segmentation & classification with vector self-attention.
Self-Attention for Point Cloud Vector Self-Attention 3D Segmentation Position Encoding Point Cloud Classification
PCT: Point Cloud Transformer
arXiv:2012.09688 (2020)
Point Cloud Transformer with offset-attention and neighbor embedding for permutation-invariant processing.
Offset-Attention Neighbor Embedding Point Cloud Transformer Global Feature Aggregation Permutation Invariance
Point-BERT: Pre-training Point Cloud Transformers
arXiv:2111.14819 (2021)
BERT-style pre-training for point cloud Transformers via masked point modeling with dVAE tokenizer.
Masked Point Modeling BERT-Style Pre-training 3D Self-Supervised Learning dVAE Tokenizer Point Cloud Transformer
Point-MAE: Masked Autoencoders for Point Cloud Pre-training
arXiv:2203.06604 (2022)
Masked autoencoders for point cloud self-supervised learning with asymmetric encoder-decoder design.
Masked Autoencoder Self-Supervised Learning Point Cloud Pre-training Asymmetric Encoder-Decoder Masked Point Reconstruction
Point-M2AE: Multi-scale Masked Autoencoders
arXiv:2205.14401 (2022)
Multi-scale masked autoencoders for hierarchical point cloud pre-training with pyramid feature learning.
Multi-Scale Masked Autoencoder Hierarchical Point Cloud Self-Supervised Pre-training Pyramid Feature Learning
2018–2023

C. Implicit 3D Surface Reconstruction

Reconstructing continuous 3D surfaces from point clouds or 2D observations via occupancy fields, signed distance functions, and neural fields. From global implicit functions (IM-NET, Occupancy Networks, DeepSDF) to local/convolutional variants (LDIF, ConvONets, LIG) and point-convolution approaches (POCO, 3DShape2VecSet). In the second thread, this is the representation-locality line: geometry becomes easier once the latent is attached to regions, grids, or points instead of being compressed into one global code.

Second thread Local geometric representation Continuous geometry Scalability
IM-NET: Learning Implicit Fields for Generative Shape Modeling
arXiv:1812.02822 (2018)
Learning implicit fields (inside/outside) for generative shape modeling, interpolation, and completion from latent codes.
Implicit Field Occupancy Classification Shape Interpolation Generative Shape Modeling
Occupancy Networks: Learning 3D Reconstruction in Function Space
arXiv:1812.03828 (2018)
Continuous 3D reconstruction in function space (occupancy classifier) — arbitrary-resolution mesh extraction.
Occupancy Function Continuous 3D Representation Neural Implicit Surface Mesh Extraction Function Space
DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation
arXiv:1901.05103 (2019)
Learning continuous signed distance functions for shape representation with auto-decoder architecture.
Signed Distance Function Continuous Shape Representation Auto-decoder Shape Completion Latent Shape Code
LDIF: Local Deep Implicit Functions for 3D Shape
arXiv:1912.06126 (2019)
Structured decomposition + local implicit decoders — high-quality 3D shape representation and assembly.
Local Deep Implicit Function Structured Decomposition Shape Assembly Implicit Surface Decoder
ConvONets: Convolutional Occupancy Networks
arXiv:2003.04618 (2020)
Translation-equivariant implicit 3D reconstruction — convolutional feature grids + occupancy decoding.
Translation Equivariance Convolutional Occupancy Network 3D Surface Reconstruction Volumetric Convolution Point Cloud Conditioning
LIG: Local Implicit Grid Representations for 3D Scenes
arXiv:2003.08981 (2020)
Part-level latent embedding in a regular grid for representing complex 3D scenes with local implicit decoders.
Local Implicit Grid Part-Level Embedding 3D Scene Representation Implicit Neural Representation Grid Feature Encoding
POCO: Point Convolution for Surface Reconstruction
arXiv:2201.01831 (2022)
Point convolution for surface reconstruction — per-point latent vectors + learned interpolation for occupancy.
Point Convolution Surface Reconstruction Per-Point Latent Vector Learned Interpolation Occupancy Prediction
3DShape2VecSet: Neural Field Representation for 3D Generation
arXiv:2301.11445 (2023)
Set-of-vectors neural field representation for diffusion-based 3D shape generation with cross-attention decoding.
Set-of-Vectors Representation Neural Field Diffusion-Based Generation Cross-Attention Decoding
2020–2021

D. CAD / B-Rep Learning & Processing

Directly operating on or generating CAD boundary representations. These methods tackle the challenge of learning from the native data structures of engineering CAD — UV-grid parameterized faces/edges (UV-Net) and topological message passing on face/edge/coedge graphs (BRepNet). In the second thread, these papers expose the target semantics any reverse pipeline must eventually satisfy: coedges, UV domains, adjacency, and CAD-valid boundary structure.

Second thread Validity constraints Native topology Engineering semantics
UV-Net: Learning from Boundary Representations
arXiv:2006.10211 (2020)
Learning from B-Rep via UV-grid parameterization + graph neural networks for CAD feature learning.
UV-Grid Parameterization B-Rep Learning CAD Feature Learning 1D/2D Convolution on B-Rep Graph Neural Network
BRepNet: Topological Message Passing on B-Rep for CAD Segmentation
arXiv:2104.00706 (2021)
Topological message passing on B-Rep faces/edges/coedges for CAD model segmentation.
Topological Message Passing B-Rep Segmentation Face/Edge/Coedge Graph CAD Model Analysis Geometric Deep Learning
2021–2024

E. 3D Shape Generation (All Representations)

Generative models for 3D shapes using various representations — set-structured VAEs (SetVAE), textured mesh GANs (GET3D), latent point diffusion (LION), unified structured latents (SLAT), and VAE benchmarking with sharp-edge sampling (Dora-VAE). In the second thread, this block supplies the generative priors and evaluation pressure that make structured completion and sampling realistic instead of purely reconstructive.

Second thread Generative priors Latent hierarchy Structure-aware evaluation
SetVAE: Hierarchical VAE for Set-Structured Data
arXiv:2103.15619 (2021)
Hierarchical VAE for set-structured data (point cloud generation) — multi-scale latent representations.
Hierarchical VAE Set-Structured Data Point Cloud Generation Latent Hierarchical Learning Permutation-Invariant Encoding
GET3D: A Generative Model of High Quality 3D Textured Shapes
arXiv:2209.11163 (2022)
Generative model of textured 3D meshes from 2D images — DMTet + differentiable rendering.
Textured 3D Mesh Generation DMTet Differentiable Rendering 3D GAN Shape and Texture Synthesis
LION: Latent Point Diffusion Models for 3D Shape Generation
arXiv:2210.06978 (2022)
Latent point diffusion models for 3D shape generation — hierarchical VAE latent space + denoising diffusion.
Latent Point Diffusion 3D Shape Generation Hierarchical VAE Denoising Diffusion Shape Latent Space
SLAT: Structured 3D Latents for Scalable 3D Generation
arXiv:2412.01506 (2024)
Structured 3D Latents — unified latent for scalable 3D generation (Radiance Fields / 3DGS / Mesh).
Structured 3D Latents Unified 3D Generation Radiance Field 3D Gaussian Splatting Mesh Generation
Dora-VAE: Sampling & Benchmarking for 3D Shape VAEs
arXiv:2412.17808 (2024)
Sampling & benchmarking for 3D shape VAEs — sharp edge importance sampling methodology.
3D Shape VAE Sharp Edge Sampling VAE Benchmarking 3D Shape Generation Surface Reconstruction Quality
2016–2020

F. Foundational / Adjacent Methods

Foundational techniques that underpin the broader research landscape: unsupervised graph learning (VGAE), interaction graph inference from dynamics (NRI), Transformer-based set prediction (DETR), and hierarchical vector graphics generation (DeepSVG). In the second thread, these are the transport mechanisms that migrate across categories: set prediction, latent graph inference, and hierarchical vector decoding.

Second thread Set / graph mechanisms Variable-count prediction Reusable abstractions
VGAE: Variational Graph Auto-Encoders
arXiv:1611.07308 (2016)
Variational Graph Auto-Encoders — foundational unsupervised learning on graphs for link prediction.
Graph Autoencoder Latent Variable Model Graph Convolution Link Prediction
NRI: Neural Relational Inference for Interacting Systems
arXiv:1802.04687 (2018)
Neural Relational Inference — inferring interaction graphs + dynamics from trajectories with VAE.
Neural Relational Inference Interaction Graph Trajectory Prediction Variational Autoencoder Graph Neural Network
DETR: End-to-End Object Detection with Transformers
arXiv:2005.12872 (2020)
End-to-End Object Detection with Transformers — set prediction, bipartite matching, no NMS or anchors.
Set Prediction Bipartite Matching Object Query End-to-End Detection Transformer
DeepSVG: Hierarchical Generative Network for Vector Graphics
arXiv:2007.11301 (2020)
Hierarchical generative network for vector graphics (2D SVG) animation and controllable generation.
SVG Generation Hierarchical Vector Graphics Path Decoder 2D Vector Representation