The movement of the ping pong ball is going to be ballistic, so you really only need to know its 3d position in 3 different locations in order to fully constrain its motion. In reality, you will probably want more. This excellent paper An Application of human-robot interaction: Development of a ping-pong playing robotic arm uses around 8-10 locations in order to fully constrain the ballistic motion, including the bounce off the table. It also has a review of past systems and approaches.
You really need to think not about the frame rate of the camera but the processing rate of the system.If you say you need 6 3d locations between both sides of the table then having more than:
$$ \frac{6 frames}{.7 sec} = 8.5 fps $$
does not really help you, but it imposes another limitation, you need to be able to capture, transfer and process images in less than:
$$ \frac{1 sec}{8.5} = 117ms $$
This usually ends up being the limitation for these kinds of systems.
Additional food for thought, syncing the camera's shutter is usually pretty important or at least sampling the images close to one another. For example with two 30FPS cameras the maximum time between two frames acquired by the computer at the "same" time is 33ms but in the time the ball will have traveled
$$ \left(\frac{2.7m}{0.7 sec} * 33ms \right) = 12.8 cm $$
Depending on how your analyzing the data this might be significant (hint it probably is) This is why some people might be using 120fps cameras reducing the max error to around 3.2cm or you might use cameras that have a global shutter clock/sync to make sure that every image captured by all your cameras are taken at the same instant. The Firefly MV camera series is an inexpensive example of cameras that can have their shutters synchronized.
TL;DR
The Kinect is probably fast enough for your purpose, it's certainly cheaper than using Point Grey cameras, so I would give it a shot, I think moving a robot arm fast and accurate enough will be a bigger challenge.