Instead of replacing with the camera transform view vector (-z), we could try to improve the plane-fitting method instead.