Room Layout Estimation Using a Machine Learning Technique

The purpose of this paper is to propose a room layout estimation pipeline from indoor point cloud data using a machine learning technique. During data pre-processing, each point cloud was aligned along the X, Y, and Z axes, and then projected along the XZ dimensions to a 2D image called a point density image; the number of points along the Y dimension (room height) that fell into the same pixel coordinate was also counted. The pixels with a high accumulation of points were typically wall areas and were colored in a higher color range index. We trained pairs of point density images and ground truth images using the U-Net model. The trained model could predict rough wall areas of unseen images. Post-processing was applied to the predicted result in order to estimate connected walls, generate a vector image, and draw the length of the walls. Because the results from the automated process always contained some errors, we created a user interface to complete the final layout with just a few modifications.