It is "stereo camera" except, instead of using two cameras at a single instance in time it uses one camera at two (or more) instances in time and uses the shifting position of the camera over time to get the multiple perspectives.

The very hard parts are tracking the camera's movement accurately enough to get valid multiple perspectives and the computer horsepower to turn the multiple images (perspectives) into 3D. Obviously, your 3D accuracy is going to be heavily dependent on the accuracy of your camera position, the resolution of the camera, and the ability to do the math in (near) real time.

One trick I've noticed in demos is that they overlay the image on the 3D model. This is really cool, but your eyes and brain will fool you into thinking the 3D model is has better resolution than it really is. If you just display the 3D model as surfaces, you will better see the limitations of the 3D model.

it's called "structure from motion", I am pretty sure Google uses ETH's code (was open source, now ETH's project website vanished), see link above (near "edit").

Actually, it is (at least partially) like the Kinect 1; it includes a PrimeSense chip. See steps 9 and 10 of the Project Tango teardown at ifixit: http://www.ifixit.com/Teardown/Project+Tango+Teardown/23835