Research map / Detection and segmentation / Open-set and few-shot detection
ArchCAD-400K: A Large-Scale CAD drawings Dataset and New Baseline for Panoptic Symbol Spotting
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.
Builds ArchCAD-400K, a 413K-chunk panoptic symbol spotting dataset of architectural CAD drawings produced with a layer-block annotation engine plus expert refinement, and proposes DPSS, a dual-pathway (image + point) model with adaptive fusion that beats prior methods.
Problem
Panoptic symbol spotting in architectural CAD drawings is limited by the small size, low diversity, and costly manual line-grained annotation of existing datasets such as FloorPlanCAD.
Approach
Automatically generate primitive-level panoptic labels from layer and block structure of standardized CAD files (refined by drafters), and fuse rendered-image features with primitive point-cloud features through a geometry-guided adaptive fusion module in a Mask2Former-style decoder.
Key results
- On ArchCAD-400K, DPSS reaches 70.6 total PQ without priors vs 60.5 for SymPointV2.
- On FloorPlanCAD without priors, DPSS reaches 86.2 PQ vs 83.2 for SymPointV2.
- Adding the adaptive fusion module raises PQ from 83.0 to 86.2 on FloorPlanCAD ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DPSS instead of SymPoint on FloorPlanCAD without priors | Panoptic symbol spotting · FloorPlanCAD (without priors) | Total PQ: 83.3 -> 86.2 | Helped |
| use DPSS instead of SymPointV2 on FloorPlanCAD without priors | Panoptic symbol spotting · FloorPlanCAD (without priors) | Stuff PQ: 49.3 -> 64.7 | Helped |
| use DPSS instead of SymPointV2 | Panoptic symbol spotting · ArchCAD-400k | Total PQ: 60.5 -> 70.6 | Helped |
| use only the primitive encoder instead of both encoder branches (ablation) | DPSS · FloorPlanCAD | PQ: 81.7 -> 86.2 | Hurt |
| use only the image encoder instead of both encoder branches (ablation) | DPSS · FloorPlanCAD | PQ: 80.9 -> 86.2 | Hurt |
| remove the image encoder and use only the primitive encoder (ablation) | DPSS · FloorPlanCAD | PQ: 81.7 -> 83.0 | Hurt |
| replace adaptive fusion with simple concatenation (ablation) | DPSS · FloorPlanCAD | PQ: 83.0 -> 86.2 | Hurt |
Surprising findings
- With layer and color priors, CADSpotting and SymPointV2 slightly outperform DPSS on FloorPlanCAD, so DPSS's advantage appears mainly without priors.
Limitations
- Cannot process an entire vector drawing in a single pass, leading to high computational overhead.
- Annotation still requires expert refinement of automated labels; non-standard drawings are excluded.
Experimental setup
Datasets: ArchCAD-400K FloorPlanCAD COCO-Stuff ScanNetV2
Benchmarks: ArchCAD-400K FloorPlanCAD
Metrics: PQ SQ RQ F1 mAP
Baselines: CADTransformer SymPoint SymPointV2 CADSpotting GAT-CADNet DeepLabv3+ DINO
Builds on: CADTransformer SymPoint SymPointV2 Mask2Former DETR
Closest papers by meaning
- Point or Line? Using Line-based Representation for Panoptic Symbol Spotting in CAD Drawings (NeurIPS 2025 · similarity 0.91)
- Symbol as Points: Panoptic Symbol Spotting via Point-based Representation (ICLR 2024 · similarity 0.90)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.85)
- FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models (ICLR 2025 · similarity 0.85)
- DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model (NeurIPS 2026 · similarity 0.85)
- Text2CAD: Generating Sequential CAD Designs from Beginner-to-Expert Level Text Prompts (NeurIPS 2024 · similarity 0.85)
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.