Training Strategies for CNN-based Models to Parse Complex Floor Plans

Abstract

A floor plan is one of the most fundamental diagrams for architectural design. Considering a large proportion of floor plans are rasterized images, we believe that parsing the rasterized images is a crucial procedure to automate architectural design. In this study, we evaluate the use of convolutional neural network (CNN) based image segmentation methods to handle floor plan parsing, instead of traditional measures such as template matching. Especially, we analyzed samples whose features are difficult for CNN-based models to learn; thus, we propose two training strategies, separate training and the use of a weighted loss function, to improve the learning accuracy for such complex samples. Experimental results demonstrate that the proposed strategies performed well for the complex samples, generating more favorable parsing output.