Research map / Point clouds / Point cloud classification
Point or Line? Using Line-based Representation for Panoptic Symbol Spotting in CAD Drawings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes VecFormer, which represents CAD vector primitives as sequences of line segments fed to a PTv3 backbone with a dual-branch Transformer decoder and a training-free Branch Fusion Refinement step, reaching 91.1 PQ on FloorPlanCAD.
Problem
Panoptic symbol spotting in vector CAD drawings requires detecting countable things and uncountable stuff, but rasterization, graph, and point-based representations are costly or lose geometric structure.
Approach
Encode each vector primitive as a sequence of line segments (line sampling) with intrinsic geometry and primitive-level pooled statistics, then apply a PTv3 backbone and dual-branch (instance and semantic) Transformer decoder, with a confidence-based fusion step (BFR) to make the panoptic output consistent.
Key results
- VecFormer reaches 91.1 PQ on FloorPlanCAD with prior information and 88.4 PQ without, outperforming the second-best methods by 1.0 and 2.2 points.
- Stuff-PQ reaches 90.4 with prior (+9.6 over SymPoint-V2) and 85.9 without (+21.2 over DPSS).
- Branch Fusion Refinement adds about 1.9 PQ (89.2 to 91.1) in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use line sampling instead of point sampling (ablation) | VecFormer · FloorPlanCAD w/o Prior | PQ: 87.8 -> 88.4 | Helped |
| use line sampling instead of point sampling (ablation) | VecFormer · FloorPlanCAD w/ Prior | PQ: 89.8 -> 91.1 | Helped |
| set sampling ratio to 0.01 rather than 0.1 (ablation) | VecFormer · FloorPlanCAD | PQ: 90.4 -> 91.1 | Helped |
| reduce sampling ratio from 0.01 to 0.005 (ablation) | VecFormer · FloorPlanCAD | PQ: 91.1 -> 90.4 | Hurt |
| add Branch Fusion Refinement (ablation) | VecFormer · FloorPlanCAD | PQ: 89.2 -> 91.1; (+1.9) | Helped |
| add Branch Fusion Refinement (ablation) | VecFormer · FloorPlanCAD | PQth: 90.4 -> 91.8; (+1.4) | Helped |
| add Branch Fusion Refinement (ablation) | VecFormer · FloorPlanCAD | PQst: 88.1 -> 90.4; (+2.3) | Helped |
| replace primitive ID with layer ID (ablation) | VecFormer · FloorPlanCAD w/ Prior | PQ: 88.4 -> 90.2 | Helped |
| add Layer Feature Enhancement (ablation) | VecFormer with layer prior · FloorPlanCAD w/ Prior | PQ: 90.2 -> 91.1 | Helped |
| use VecFormer instead of the strongest reported baseline | FloorPlanCAD panoptic symbol spotting · FloorPlanCAD w/o Prior | PQst: +21.2 | Helped |
| use VecFormer instead of SymPoint-V2 | FloorPlanCAD panoptic symbol spotting · FloorPlanCAD w/ Prior | PQst: +9.6 | Helped |
| use VecFormer instead of SymPoint-V2 | FloorPlanCAD primitive-level semantic prediction · FloorPlanCAD | F1: +4.3 | Helped |
Surprising findings
- Replacing layer IDs with primitive IDs as the z-coordinate gives 88.4 PQ without prior, close to the 91.1 with layer priors, indicating reduced dependence on layer annotations.
Limitations
- Uniform line sampling may underperform in regions of high geometric complexity.
- Evaluated only on FloorPlanCAD.
Experimental setup
Datasets: FloorPlanCAD
Benchmarks: FloorPlanCAD
Metrics: PQ PQ-thing PQ-stuff F1 wF1
Baselines: PanCADNet CADTransformer GAT-CADNet SymPoint SymPoint-V2 CADSpotting DPSS
Builds on: OneFormer3D Point Transformer V3 SymPoint-V2 Mask2Former
Closest papers by meaning
- Symbol as Points: Panoptic Symbol Spotting via Point-based Representation (ICLR 2024 · similarity 0.92)
- ArchCAD-400K: A Large-Scale CAD drawings Dataset and New Baseline for Panoptic Symbol Spotting (NeurIPS 2025 · similarity 0.91)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.86)
- DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model (NeurIPS 2026 · similarity 0.84)
- CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention (CVPR 2024 · similarity 0.83)
- CADTalk: An Algorithm and Benchmark for Semantic Commenting of CAD Programs (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.