<p>In this study, we introduce a transformer-based floor plate plan architecture, FPP-Former, as an end-to-end framework for the semantic and structural reconstruction of large-scale floor plate plans (FPPs). Unlike traditional pixel-wise semantic segmentation methods, our approach employs a patch-based network to extract semantically meaningful masks from FPP images that capture entire building levels. By integrating these semantic masks with differentiable rasterization techniques, the framework extracts contour polygons that represent various architectural elements, including rooms, doors, and windows. Subsequently, a downsampler is employed to simultaneously reduce redundant polygon vertices and preserve the original geometric integrity of each polygon. By consolidating semantically labeled polygons from all constituent patches, our framework achieves efficient and accurate reconstruction of FPP images at arbitrary resolutions. To advance research in this field, we also introduce an FPP dataset, FPP-Set, as a comprehensive dataset comprising high-resolution images sourced from legally authorized residential computer-aided design documents. With an average resolution of 70 million pixels, the FPP-Set provides a comprehensive and high-fidelity representation of residential FPP and enables a detailed examination of floor plan reconstruction methods. Experimental results on FPP-Set and existing benchmarks highlight the exceptional capability of the FPP-Former in delivering accurate, reliable reconstructions of complex and challenging FPP images.</p>