Machine Learning-Based Prediction of Floor Plan Symbols for Façade Design Alternative Generation

This dataset contains 5000 typical residential floor plans in image format along with their corresponding annotation files. The plans were generated by considering a fixed architectural program while exploring different spatial arrangements using Grasshopper and Magnetizing plugins . Using this dataset, a YOLO-based model was trained for floor plan image analysis and 3D model generation to produce diverse façade alternatives. The results demonstrate remarkable performance, with mAP, precision, and recall of 0.993, 0.997, and 0.99, respectively, leading to high 3D model accuracy (98.99% topology consistency) and window layout alternative generation accuracy of 99.3%.