Now that 3D sensors have become widely affordable, there will be an increasing need to develop ways for robots to interact with 3D environments instead of 2D approximations. 3D mapping is going to be one of the more important uses for these sensors. At a practical level, the question quickly becomes, how are you going to deal with all of that data? Octrees are one solution.
This visualization by Jose Luis Blanco Claraco uses MRPT for efficiently rendering pointclouds of 11 million points. Which in terms of scale is only slightly more that a second worth of data from the Kinect at 640x480x30fps.
OctoMap is also available for ROS.