Segmentation of Spatial and Geometric Information from Floorplans using CNN Model

Abstract

In automated document analysis, floorplans are a concern for several years and algorithmic approaches have been used until recently. This problem has also improved output with the emergence of Convolutionary neural networks (CNN). In this study, it is the task to retrieve space and geometric data from planes as accurately as possible and the bulk of the information is extracted from a plane image by means of instance segments, such as the Cascade Mask R-CNN. A new way of using keypoint CNN is presented in order to supplement the segmentation, so that precision corner positions can be identified. Then they are coupled to the resulting segmentation in a post-processing stage. With an average IoU of 72.7% compared to 57.5%, the resulting segmentation scores surpass the existing benchmark for CubiCasa5k floorplan datasets. In addition, the mean IoU is increased in almost every class for individual classes. Cascade masks R-CNN have also shown to be more appropriate for this mission than R-CNN masks.