Research map / Point clouds / Point cloud classification
Symbol as Points: Panoptic Symbol Spotting via Point-based Representation
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.
Treats CAD vector-graphic primitives as 2D point clouds and applies a Point Transformer backbone with a Mask2Former-style spotting head, plus connection-aware attention, contrastive connection learning and KNN mask downsampling, for panoptic symbol spotting on FloorPlanCAD.
Problem
Panoptic symbol spotting in CAD vector drawings: detecting and parsing both countable instances (doors, windows, furniture) and uncountable stuff (walls, railings). Prior methods either rasterize drawings (costly) or use graph neural networks over primitives.
Approach
Represent each graphical primitive as a point (position plus primitive features) and treat the drawing as a point cloud; add attention over locally connected primitives (ACM), a contrastive loss over connections (CCL), and KNN-based interpolation to downsample primitive-level attention masks in a Mask2Former-like head.
Key results
- On FloorPlanCAD panoptic symbol spotting, SymPoint (1000 epochs) reaches 83.3 PQ, 91.1 RQ and 91.4 SQ, versus GAT-CADNet's 73.7 PQ.
- On semantic symbol spotting, SymPoint achieves 86.8 F1 and 85.5 wF1 versus GAT-CADNet's 85.0 F1 and 82.3 wF1.
- Ablation: KNN interpolation for mask downsampling lifts PQ from 74.3 to 77.3 at 300 epochs; the full 300-epoch model reaches 79.6 PQ at 66 ms inference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SymPoint instead of GAT-CADNet | FloorPlanCAD semantic symbol spotting · FloorPlanCAD | F1: 85.0 -> 86.8 | Helped |
| Train SymPoint for 1000 rather than 300 epochs | SymPoint · FloorPlanCAD panoptic symbol spotting | PQ: 79.6 -> 83.3 | Helped |
| Add ACM alone (ablation) | SymPoint baseline · FloorPlanCAD panoptic symbol spotting, 300 epochs | PQ: 73.1 -> 72.6 | Hurt |
| Add CCL alone (ablation) | SymPoint baseline · FloorPlanCAD panoptic symbol spotting, 300 epochs | PQ: 73.1 -> 73.5 | Helped |
| Add ACM together with CCL (ablation) | SymPoint baseline · FloorPlanCAD panoptic symbol spotting, 300 epochs | PQ: 73.1 -> 74.3 | Helped |
| Add KNN interpolation to ACM and CCL (ablation) | SymPoint · FloorPlanCAD panoptic symbol spotting, 300 epochs | PQ: 74.3 -> 77.3 | Helped |
| Replace linear mask downsampling with KNN average pooling (ablation) | SymPoint · FloorPlanCAD panoptic symbol spotting, 300 epochs | PQ: 74.3 -> 75.9 | Helped |
| Replace linear mask downsampling with KNN max pooling (ablation) | SymPoint · FloorPlanCAD panoptic symbol spotting, 300 epochs | PQ: 74.3 -> 77.0 | Helped |
| Use KNN interpolation instead of linear mask downsampling (ablation) | SymPoint · FloorPlanCAD panoptic symbol spotting, 300 epochs | PQ: 74.3 -> 77.3 | Helped |
| Increase backbone size from 1x to 1.5x (ablation) | SymPoint · FloorPlanCAD panoptic symbol spotting, architecture ablation | PQ: 67.1 -> 73.3 | Helped |
| Increase backbone size from 1.5x to 2x (ablation) | SymPoint · FloorPlanCAD panoptic symbol spotting, architecture ablation | PQ: 73.3 -> 77.3 | Helped |
| Increase query count from 300 to 500 (ablation) | SymPoint with 2x backbone · FloorPlanCAD panoptic symbol spotting, architecture ablation | PQ: 77.3 -> 77.9 | Helped |
Surprising findings
- Crudely enlarging the attention radius without connection information did not improve performance, and ACM alone slightly reduced PQ (73.1 to 72.6).
- Applying CCL alone gave only marginal gains, while ACM combined with CCL gave a larger improvement.
Limitations
- Requires long training (1000 epochs) to reach best performance.
- Private dataset is not public, limiting reproducibility of the additional experiments.
- Noisy connections between locally connected primitives from different categories can hurt performance.
Experimental setup
Datasets: FloorPlanCAD SESYD-floorplans private floorplan dataset
Benchmarks: FloorPlanCAD SESYD-floorplans
Metrics: PQ RQ SQ F1 wF1 AP50 AP75 mAP
Baselines: PanCADNet CADTransformer GAT-CADNet PointT Cluster FasterRCNN YOLOv3 FCOS DINO YOLaT RendNet
Builds on: Point Transformer Mask2Former InfoNCE
Closest papers by meaning
- Point or Line? Using Line-based Representation for Panoptic Symbol Spotting in CAD Drawings (NeurIPS 2025 · similarity 0.92)
- ArchCAD-400K: A Large-Scale CAD drawings Dataset and New Baseline for Panoptic Symbol Spotting (NeurIPS 2025 · similarity 0.90)
- Hierarchical Process Reward Models are Symbolic Vision Learners (CVPR 2026 · similarity 0.81)
- Primitive Vision: Improving Diagram Understanding in MLLMs (ICML 2025 · similarity 0.81)
- SSAN: A Symbol Spatial-Aware Network for Handwritten Mathematical Expression Recognition (AAAI 2025 · similarity 0.81)
- Detect Anything in Graphic Design: Element-Level Rewards for Autoregressive Detection (NeurIPS 2026 · similarity 0.80)
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.