Abstract

We introduce CADSpotting, an effective method for panoptic symbol spotting inlarge-scale architectural CAD drawings. Existing approaches struggle withsymbol diversity, scale variations, and overlapping elements in CAD designs.CADSpotting overcomes these challenges by representing primitives throughdensely sampled points with attributes like coordinates and colors, using aunified 3D point cloud model for robust feature learning. To enable accuratesegmentation in large, complex drawings, we further propose a novel SlidingWindow Aggregation (SWA) technique, combining weighted voting and Non-MaximumSuppression (NMS). Moreover, we introduce LS-CAD, a new large-scale CAD datasetto support our experiments, with each floorplan covering around 1,000 squaremeters, significantly larger than previous benchmarks. Experiments onFloorPlanCAD and LS-CAD datasets show that CADSpotting significantlyoutperforms existing methods. We also demonstrate its practical value throughautomating parametric 3D reconstruction, enabling interior modeling directlyfrom raw CAD inputs.