Further, features of a user's face can be extracted for each captured frame to determine the location of a user's eyes and mouth, for example 1512. The change in position of these features between subsequent image frames can be used to determine what tracking method will be used for that particular frame. For example, the change in optical flow of a user's eyes can be calculated for a current and previous image frame pair. In one example, if this change is less than a first amount or threshold, then the position of the user's eyes has only slightly changed relative to their position in the previous frame. Since this change is small, the user's current eye position can be reasonable estimated as its location in the previous frame, as if the user hasn't moved. In another example, if this change is between the first threshold and a second threshold, a single point tracking algorithm can be used to track the user's eyes between these two frames in order to reduce jitteriness. If, however, this change in optical flow is greater than the second threshold, the current position of the user's eyes can be used. In this instance, the tracking output will appear quite jittery, however, since the change in eye position is so great (i.e., greater than the second threshold) the user has moved quickly or abruptly and, thus, an abrupt change, in this instance, would not only be acceptable, it would likely be expected. Once the current location of the eyes, in this example, is determined for each image captured by each camera, stereo disparity between the eyes in each of these images is determined. The stereo disparity is then used to determine a z-depth for the eyes, by calculating a distance between the eyes and the computing device, in order to determine a three-dimensional position (x, y, z) of the eyes relative to the computing device. Accordingly, a point relative to one or more features of the user's face can be tracked in three-dimensions (x, y, z) in order to provide a smoother and more accurate location of the user relative to a computing device. Additionally, once these features have been identified, various heuristics can be applied to reject false positive user detections 1516.