I am confronted with the same problem, if i've understood the preceding post correctly. The images below shows the walls, as detected by apple room plan, over a heat-map (reflecting point densities) of the corresponding ARSession's point cloud projected on the floor plane. To be clear, the point cloud is constructed from depth-frames recorded during the room capture session. Also, as the point cloud is not fragmented, we can rule out that a loss of world tracking causes this behavior. IMHO, this behavior is a bug as it destroys the correspondence of the estimated surfaces with related RGB and depth data and could be avoided (?). Hopefully this gets fixed soon, cheers c.