CubiCasa5k-ScanShift

CubiCasa5k-ScanShift is a derived floor plan dataset release for evaluating object detection robustness under scan-induced domain shift in the study "Why Floor Plan Object Detection Breaks Under Real Scans: A Controlled Study of Robustness Limits." The archive contains 4,996 validated layouts with aligned annotations and paired image variants. For each retained layout, the release includes graph.json, the clean raster model_baked.png, the scanned raster F1_scaled.png, and two ROI-controlled variants: four_final_variants/03_scan_inside_boxes.png and four_final_variants/04_svg_clean_plus_scan_outside.png. The release also includes corrected SVG-derived renderings and visualization outputs where preparation fixes were required for consistency and reproducibility. The distributed split files were filtered so that every listed layout contains the complete required file set. The retained split counts are 4,197 training layouts, 400 validation layouts, and 399 test layouts. A machine-readable staging report is included as DATASET_STAGE_SUMMARY.json. This release supports reproducible evaluation of clean-to-scan transfer, controlled foreground/background isolation, and robustness analysis for floor plan object detection.