One common use for ORB, is in tracking and identifying objects in real time video streams. In this case, we compute the ORB descriptors for any images or objects we want to detect, before seeing a video stream and save those descriptors. Then, for each frame in an incoming video stream, we calculate ORB descriptors and use a matching function to compare the key points in the current video frame with the saved descriptors. For any object descriptor that we’ve saved. If we find that the matching function returns a number of matches above some match threshold, we can conclude that the object is in the frame. The mesh threshold is a free parameter that you can set. For example, if the ORB descriptor for a particular object has 100 key points, then you could set the threshold to be 35 percent, 50 percent, or 90 percent of the number of key points for that particular object. If you set the threshold to 35 percent, then that means that at least 35 key points out of the 100 that describe that object, must match in order to say that the object is in the frame. All these steps can be done in near real time because ORBs binary descriptors, are extremely fast to compute and compare. The ORB algorithm works best when you want to detect objects that have a lot of consistent features that are not affected by the background of an image. For example, ORB works well for facial detection, because faces have a lot of features such as the corner of the eyes and the mouth, that don’t appear to change no matter where a person is. These features are consistent from image to image. However, ORB does not work so well when attempting to do more general object recognition. Say pedestrian detection in images, in which the shape and features of a person’s body vary depending on clothing and movement. For this type of general object recognition, other algorithms work much better. In the next section, we’ll learn about methods that can be used to do more general object recognition.