Ensuring a proper, easy, and convenient interface between human and machine has been one of the most challenging issues since the first machines were developed. In the beginning, when simple machines responded to direct human hand interaction with different levers, things were rather easy.

For example, let’s consider pincers: human error, like using too little force or the wrong handle, did not usually lead to serious consequences. As machines become more and more complicated, and pressing one button can lead to a series of actions, this is not so obvious. The first simple human-computer interfaces based on keyboards and monitors guaranteed the possibility to correct what was wrongly written after the prompt sign, or when “enter” was pressed, to see the system response. A wrong command led to a syntax error or question presented on the screen: are you sure? When graphical user interfaces (GUIs) became more popular and the mouse was used, our senses learned a new kind of interface; however, for many, it took some time to coordinate the hand and mouse pointer on the screen without looking at the hand controlling the mouse. How many wrong movements and wrong clicks were the result? Next came the use of touch interfaces, especially in mobile devices. Note: many people are afraid of touching a “wrong” button.

Now we are expected to use voice or gesture-based interfaces, but such sophisticated solutions are always related to issues of reliability and gesture recognition. Several challenges are still open, but the most important is: “How can one discriminate everyday movement from intentional movement?” This leads to problems of false positives (“systems respond without the user intending to invoke a command”), false negatives (a system does not respond despite the right gesture command), and “recognizer accuracy ... expressed in terms of recall and precision.” Such cases can lead to very costly actions. On the other hand, if a system is too sensitive, every movement can fire actions and thus lead to the “Midas touch effect.”

As explained in the paper, “the typical approach to balancing precision and recall [is based on] thresholds, which ensure that only input that is sufficiently close to an individual category is recognized.” However, the authors propose a modified recognition strategy called bi-level thresholding, which looks at interaction as a sequence of user actions. If a gesture is slightly incorrect, the system examines another input. If this one is similar (a near miss) the gesture is recognized. Based on experiments with five gestures, around two-thirds of the attempts were recognized with the bi-level threshold model. The paper includes experimental data based on a bi-level thresholding state diagram (figure 8) with background related to a bi-level threshold recognizer (figure 1).

Finally, the authors state: if a user’s gesture is a near miss, such a situation also provides “valuable information which can be used to enhance the perceived reliability of recognition-based interactions.”