Why floor plan object detection breaks under real scans: a controlled study of robustness limits

Robust object detection in architectural floor plans is essential for construction automation, yet real scans alter geometric depiction. This paper presents a controlled study based on CubiCasa5k, using paired CAD-derived and scanned raster images with identical layouts and annotations to isolate appearance. Three-seed RT-DETR experiments and Region-of-Interest composites separate object-region depiction shifts from background texture and clutter. Results show strongly class- and scale-dependent degradation: AP 75 and AP@0.85 deteriorate substantially more than AP 50 . Altered geometric depiction within object regions is the primary source of loss, whereas scanned backgrounds produce only a moderate decline. Single-run Faster R-CNN and RetinaNet checks show the same qualitative localization-loss signature without supporting detector ranking. Rather than introducing a new detector, this paper establishes a diagnostic benchmark and causal framework for robust detection of thin and small structures. The paired dataset variants, verification scripts, and trained checkpoints are publicly released. • Paired CubiCasa5k layouts isolate scan-induced depiction shifts. • ROI variants separate geometry depiction from scanned background effects. • Scan artifacts harm strict localization more than coarse detection. • Geometry depiction shifts dominate the performance loss. • Detector-family checks support the pattern; data and checkpoints are released.