The markerless outdoor AR system described in [13] treats walls,
building facades and ground as natural plane surfaces,recovering the camera position by estimating the reverse projection from the 3D world to the image using a Random Forest classifier.
It requires a training step that takes less than 5 minutes, using a set of 1,000 synthetically generated new examples. While it is a good algorithm for outdoor environments, it is too computationally expensive for a system that will be run in a controlled environment.