P‐2.23: Deep Learning‐based System for Processing Complex Floorplans

Abstract

With the increasing number of augmented reality apps for houses in recent years, home modeling is essential to complete a 3D reconstruction via identifying the primary features of the house based on a 2D floorplan. Due to the dispersed wall arrangement in 2D floor layouts and the abundant interference information surrounding varied thicknesses, existing segmentation methods mainly rely on image morphology or use deep learning models in other fields like Unet. However, these schemes do not solve poor robust performance problems. In this paper, we propose an Reflect Strip Pooling Unet (RSP-Unet) to enhance the segmentation capabilities of the network for strip features. Specifically, we utilize reflect strip pooling to replace the maximum pooling step and reduce feature loss during the downsampling in the Unet network. More importantly, the proposed module is also integrated with the SE (Squeeze-and- Excitation) mechanism to interact with input from several channels, lessen model overfitting, and increase model robustness. Finally, our extensive experience shows that the results on the self-built floorplan dataset demonstrate that the mean Intersection Over Union(mIOU) is increased by 8.34% and the Dice coefficient is increased by 8.78% compared with the original Unet model.