ArchCAD-400k: A Large-Scale CAD drawings Dataset and New Baseline for Panoptic Symbol Spotting

Ruifeng Luo2,1*, Zhengjie Liu1,5*, Tianxiao Cheng2, Jie Wang2, Tongjie Wang2, Xingguang Wei3,6, Haomin Wang3,7, Yanpeng Li2, Fu Chai1,5, Fei Cheng1, Shenglong Ye3, Wenhai Wang3, Yanting Zhang8, Yu Qiao3, Hongjie Zhang3†, Xianzhong Zhao1,4†,
1Tongji University
2Arcplus Group East China Architectural Design & Research Institute Co., Ltd.
3Shanghai AI Laboratory
4Shanghai Qi Zhi Institute
5Shanghai Innovation Institute
6University of Science and Technology of China
7Shanghai Jiao Tong University
8Donghua University
(* denotes equal contribution, † denotes the corresponding author)
arXiv Code Dataset
News
  • 2025.10.16: Source code of the DPSS framework is released !
  • 2025.10.16: ArchCAD dataset is open for download !
  • 2025.10.9: Our project homepage is launched !

🌟 Contributions

  • 🚀 Data Engine: A structure-aware annotation engine that reduces labeling cost by over 50x using CAD layers and blocks.
  • 📦 Dataset: ArchCAD-400k, a large-scale dataset with 413K annotations across 27 categories and diverse buildings.
  • 📈 Model: DPSS, a dual-pathway framework achieving state-of-the-art results with strong scalability.

🔍 Background

  • Problem: Panoptic symbol spotting aims to interpret CAD drawings by assigning semantic and instance labels to graphical primitives.
  • Value: Accurate primitive-level perception enables scalable BIM modeling and intelligent CAD understanding.
  • Challenges: Annotating CAD drawings is highly labor-intensive, limiting dataset size and model generalization.
ArchCAD

Illustration of panoptic symbol spotting in CAD drawings

🚀 Contribution 1 — Data Engine

  • Standardized Selection: 5,538 CAD drawings with consistent layer-block structures from 11,917 industry-grade samples.
  • Structure-Aware Automated Labeling: Leveraging inherent layer and block hierarchy to generate semantic and instance-level annotations at scale.
  • Expert-Guided Refinement: Human experts refine annotations directly in the vector space, ensuring high accuracy while avoiding rasterization artifacts.
ArchCAD

Overall pipeline of the annotation process

📦 Contribution 2 — ArchCAD Dataset

  • Broader Diversity: Covers a wide range of real-world building types beyond residential.
  • Larger Scale: Includes 5,500+ full drawings with an average area of 11,000 m².
  • Richer Semantics: Annotations across 27 categories, including structural, non-structural, and symbols.
ArchCAD

Comparison of ArchCAD to existing CAD drawing datasets.

ArchCAD

Distribution of ArchCAD, including (a) drawing area, (b) building types, and (c) symbol categories.

img
img
img

A glimpse of ArchCAD-400k samples

📈 Contribution 3 — DPSS framework

  • Novel Framework: Proposed DPSS, a panoptic symbol spotting framework tailored for CAD drawings.
  • State-of-the-Art Performance: Achieved new benchmarks on FloorPlanCAD and ArchCAD-400k, surpassing the second-best method by 3% and 10%, respectively.
  • Robust and Scalable: Demonstrated exceptional accuracy, robustness, and scalability across diverse CAD datasets.
ArchCAD

Overview of DPSS framework

img
img
img

Quantitive & Qualitative results

BibTeX

@article{luo2025archcad,
  title={ArchCAD-400K: An Open Large-Scale Architectural CAD Dataset and New Baseline for Panoptic Symbol Spotting},
  author={Luo, Ruifeng and Liu, Zhengjie and Cheng, Tianxiao and Wang, Jie and Wang, Tongjie and Wei, Xingguang and Wang, Haomin and Li, YanPeng and Chai, Fu and Cheng, Fei and others},
  journal={arXiv preprint arXiv:2503.22346},
  year={2025}
}