Did you have a look at the NARF section in the PCL tutorials?
Also, for such a tiny object, the Kinect's resolution is not high enough (for use with NARF, or anything else - there are simply not enough points on the object). Did you try visualize the Kinect point cloud in RViz, and had a look at it? If a human can't separate the object from the background noise, then forget about automated object recognition.
What other properties does the object have? Does it have a salient color? If yes, you would probably do much better with a high-resolution camera and some color blob detection. You mentioned trying a HSI threshold, why didn't it work?