Reproduce “MSD: A Benchmark Dataset for Floor Plan Generation of Building Complexes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Casper van Engelenburg, Fatemeh Mostafavi, Emanuel Kuhn, Yuntae Jeon, Michael Franzen, Matthias Standfest, Jan van Gemert, Seyran Khademi
Researchers from TU Delft introduced Modified Swiss Dwellings (MSD), the first large-scale benchmark dataset specifically for generating complex, multi-apartment floor plans, which features non-Manhattan room shapes and structural information. Benchmarking revealed that existing state-of-the-art models achieved significantly low MIoU scores (10.9% to 42.4%) on MSD, highlighting the dataset's challenging complexity and current limitations in AI-driven architectural design.
Key insights
- MSD sets a new, higher standard for complexity and realism, featuring multi-apartment dwellings, a significant proportion of non-Manhattan rooms, higher average corners per room (8.68 vs. 4.54-5.04), and precise structural and topological information.
- State-of-the-art floor plan generation models exhibit a substantial performance drop on MSD (MIoU 10.9% to 42.4%), indicating they are not yet equipped to handle the dataset's multi-faceted complexity effectively.
- While models struggle with geometric precision, diffusion-based approaches (MHD) show reasonable capabilities in maintaining topological connectivity (Graph Compatibility 71.1% to 87.1%), suggesting a disconnect between high-level layout understanding and low-level geometric accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.