CADSpotting: Robust Panoptic Symbol Spotting on Large-Scale CAD Drawings

Fuyi Yang1,2, Jiazuo Mu1, Mingqian Zhang1,2, Yanshun Zhang2, Junxiong Zhang1,2, Lan Xu*,1, Yingliang Zhang*,2
1ShanghaiTech University 2DGene Digital Technology *Corresponding authors
Paper arXiv Code
CADSpotting segments architectural CAD symbols and supports automated 3D interior reconstruction.
CADSpotting accurately identifies and segments symbols in CAD drawings. Dense point sampling and sliding window aggregation enable robust panoptic symbol spotting on large-scale drawings, while the resulting semantic information supports automated parametric reconstruction of architectural 3D interiors.

Abstract

We introduce CADSpotting, an effective method for panoptic symbol spotting in large-scale architectural CAD drawings. Existing approaches often struggle with symbol diversity, scale variations, and overlapping elements in CAD designs, and typically rely on additional features (e.g., primitive types or graphical layers) to improve performance. CADSpotting addresses these challenges with a primitive-agnostic, coordinate-only dense sampling representation: each CAD primitive is converted into densely sampled 2D points embedded in 3D with z = 0, and a point-cloud backbone learns features without explicit primitive-type or layer attributes. To enable accurate segmentation in large drawings, we further use sliding window aggregation (SWA), which combines weighted voting and sparse non-maximum suppression (NMS) to merge local predictions across overlapping windows. Moreover, we introduce LS-CAD, a large-scale dataset comprising 45 finely annotated complete floorplans, each covering approximately 1,000 m2 or more. LS-CAD complements existing large-quantity CAD datasets by focusing on physically extensive real-world layouts with thousands to tens of thousands of primitives per drawing. Experiments on FloorPlanCAD and LS-CAD demonstrate that CADSpotting achieves strong performance compared to existing methods. We also showcase its practical value in enabling automated parametric 3D interior reconstruction directly from raw CAD inputs.

Pipeline Overview

The CADSpotting pipeline: dense point sampling, PTv3 feature extraction, primitive pooling, and panoptic prediction.
Given a CAD drawing, CADSpotting densely samples points along graphic primitives to construct a point-cloud representation using spatial coordinates only. PTv3 extracts geometric features, primitive mixed pooling aggregates point features, and a streamlined Transformer decoder performs panoptic symbol spotting.

FloorPlanCAD Results

Qualitative comparison of CADSpotting and prior methods on FloorPlanCAD.
CADSpotting produces accurate panoptic symbol spotting in challenging scenes containing walls, complex or overlapping symbols, and uncommon furniture symbols. It achieves the best F1 and weighted F1 scores for semantic spotting, and the best AP75 and mAP for instance spotting among the compared methods.

LS-CAD and Large-Scale Results

LS-CAD contains 45 complete floorplans from extensive buildings such as campuses and office complexes. Each drawing covers at least 1,000 m2, with representative samples containing approximately 2,900 to more than 28,000 primitives. The dataset follows the fine-grained FloorPlanCAD annotation standard and is available from the corresponding author upon reasonable request.

Semantic and panoptic CADSpotting results on a large-scale LS-CAD office drawing.
Semantic and panoptic spotting using SWA on a large-scale LS-CAD office drawing with 28,193 primitives and 1,185 instances. The side panels show close-up comparisons between ground truth and predictions.
Comparison of ground truth, CADSpotting with SWA, CADSpotting with block partitioning, and SymPoint-V2 with block partitioning.
Comparison on a large-scale LS-CAD office drawing: ground truth, CADSpotting with SWA, CADSpotting with block partitioning (BP), and SymPoint-V2 with BP. CADSpotting with SWA reaches 75.5 PQ, 93.3 RQ, 93.5 F1, and 57.5 mAP on LS-CAD.

Automated 3D Interior Reconstruction

Raw architectural CAD input.
Raw CAD input
Semantic and instance prediction from CADSpotting.
CADSpotting prediction
Automatically reconstructed 3D interior model.
Automated reconstruction

CADSpotting's instance-level segmentation and primitive positions provide spatial parameters for walls, doors, and windows. Door positions, orientations, and pivot points are recovered from geometric relationships between arcs and lines; window positions are derived from their instances; and wall contours are extracted from closed polygons formed by adjacent wall primitives. These parameters drive a Blender-based parametric reconstruction pipeline that uses reusable wall, door, and window assets to generate structurally accurate 3D interiors from raw CAD inputs within minutes.

BibTeX

@article{yang2024cadspotting,
  title   = {CADSpotting: Robust Panoptic Symbol Spotting on Large-Scale CAD Drawings},
  author  = {Yang, Fuyi and Mu, Jiazuo and Zhang, Mingqian and Zhang, Yanshun and
             Zhang, Junxiong and Xu, Lan and Zhang, Yingliang},
  journal = {arXiv preprint arXiv:2412.07377},
  year    = {2024}
}