You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
To request access to the ArchCAD dataset, please complete the following form.
Access is restricted to non-commercial use.
All requests will be reviewed, and approval may take up to 3 business days.
Log in or Sign Up to review the conditions and access this dataset content.
ποΈ ArchCAD
A Multimodal CAD Dataset for Vector Data Understanding
400K samples Β· 5 strictly aligned data modalities Β· A foundational dataset for AI models to deeply understand engineering drawings
π Table of Contents
- What is ArchCAD?
- Key Features
- Dataset Structure
- Baseline Model: DPSS
- Potential Applications
- Citation
π What is ArchCAD?
Large language models (LLMs) have struggled to directly interpret and utilize the information embedded in CAD drawings, creating an industry bottleneck. ArchCAD addresses this challenge by providing a large-scale, high-quality CAD dataset built on a novel primitive-level annotation scheme. It serves as the core corpus for training sophisticated AI models capable of a deep, multi-layered understanding of vector-based engineering drawings.
The dataset is constructed using a "drawing slicing" method, which generates multiple, strictly aligned data modalities for each engineering sub-drawing.
β¨ Key Features
- π§ Rich Multimodality: Provides five types of AI training data for each sample: rasterized images, SVG, JSON, question-answer pairs, and point clouds.
- ποΈ Primitive-Level Annotation: Drawings contain multiple layers of semantic information, with detailed semantic and instance-level labels for each geometric primitive.
- 𧩠Strict Data Alignment: All five data modalities are strictly aligned for each drawing sub-section, enabling robust multimodal training.
- π Enables New Capabilities: Facilitates a range of downstream tasks, including high-fidelity vector symbol spotting, multimodal Q&A, raster-to-vector reconstruction, and instruction-based drawing generation.
𧩠Dataset Structure
Data Modalities
ArchCAD provides 5 types of data for each sample:
- Rasterized Images: Standard pixel-based representations of the CAD drawings.
- SVG Data: Vector graphics format where each geometric primitive is annotated with semantic and instance labels.
<svg> <path d="M x1, y1 L x2, y2" semantic="single_door" instance="single_door_23" stroke-width=1> <circle cx="x1" cy="y1" r="r1" semantic="single_door" instance="single_door_23" stroke-width=1> ... </svg> - JSON Data: A structured format detailing primitives, coordinates, types, and labels.
[ { "type": "LINE", "linetype": "Continuous", "start": [x1, y1], "end": [ x2, y2 ], "rgb": [ 0, 0, 0 ], "semantic": "single_door", "instance": "single_door_23", }, { "type": "CIRCLE", "center": [x1, y1], "radius": r1, "semantic": "single_door", "instance": "single_door_23", }, ... ] - Question-Answer Pairs: Text-based pairs for training models to answer queries about drawing contents.
- Q: How many doors are in the drawing?
- A: There are 11 doors in the drawing, including 7 single doors, 1 double door, 1 mother-son door, and 2 other doors.
- Q: What are the coordinates of the staircase in this drawing?
- A: The vertex coordinates of the rectangular bounding box for the staircase are: (0, 624), (104, 624), (104, 827), (0, 827).
- Point Cloud Data: A set of data points in space representing the geometry of the drawing.
Annotation Scheme
CAD drawings possess multiple layers of semantic information. ArchCAD captures this with a hierarchy of Drawing β Instance β Primitive. Each primitive (e.g., a line, an arc) is annotated with both a semanticId (what it is, e.g., "door") and an instanceId (which specific object it belongs to, e.g., "door_23").
π Baseline Model: DPSS
With the dataset, a specialized vector recognition model named DPSS was proposed for engineering drawings. The model architecture combines a Point Transformer backbone for primitive sets and a CNN backbone for raster images, using an adaptive fusion module to merge features.
DPSS achieves state-of-the-art (SOTA) performance on the panoptic symbol spotting task, outperforming previous methods.
| Method | Total (PQ/SQ/RQ) | Thing (PQ/SQ/RQ) | Stuff (PQ/SQ/RQ) |
|---|---|---|---|
| CADTransformer[12] | 60.0 / 89.7 / 66.9 | 52.5 / 83.6 / 62.7 | 70.1 / 96.7 / 72.5 |
| SymPoint[14] | 47.6 / 86.1 / 55.3 | 51.4 / 91.9 / 55.9 | 39.9 / 73.9 / 54.0 |
| SymPointV2[13] | 60.5 / 88.0 / 68.8 | 62.4 / 91.7 / 68.1 | 52.8 / 73.7 / 71.7 |
| DPSS | 70.6 / 90.2 / 78.2 | 65.6 / 92.4 / 70.9 | 77.6 / 87.8 / 88.4 |
π Potential Applications
The richness of the ArchCAD dataset enables several innovative AI applications for the AEC industry.
1. Multimodal Drawing Q&A
By fine-tuning multimodal large models (like InternVL) on the Q&A data, models can accurately respond to user commands to identify and locate key components like doors, windows, and stairs within a drawing.
2. Raster-to-Vector Reconstruction
Models can be trained to recognize and predict the feature points of primitives, allowing for the efficient reverse engineering of raster images (like scans) into structured, editable vector drawings.
3. Instruction-based Vector Drawing Generation & Editing
Using a Vision Language Model (VLM) architecture, models can be trained to generate high-precision vector graphics from scratch based on a raster image prompt or edit existing vector drawings based on natural language instructions (e.g., "Change the single door to a double door").
π Citation
If you use ArchCAD in your work, please cite the following paper:
@article{Luo2025ArchCAD,
title={ArchCAD-400K: An Open Large-Scale Architectural CAD Dataset and New Baseline for Panoptic Symbol Spotting},
author={Luo R, Liu Z, Cheng T, et al.},
journal={arXiv preprint arXiv:2503.22346},
year={2025}
}
- Downloads last month
- 11