5.4 Main Problems of Eye-Gaze Interfacing

While eye-gaze input might be fast, because of the close connection between the point of interest and the direction of gaze, it has some inherent problems, due to its subconscious nature:

How should the tracking data be treated?

The direction of gaze often pinpoints the object of interest, but not always, cf. section 3.4. Furthermore, several fixations in a small area might indicate that the viewer is interested in a single object at that location. Thus the eye-gaze interface designers are challenged with the problem of making the system react to what the user thinks she is looking at, not what she is actually looking at (Jacob 1993, p. 165).

To this end, Just & Carpenter (1980) used a technique of aggregating several fixations on the same object (word) into one unit they called a gaze. This technique has also been adopted by Jacob (1993), who preprocess the eye-gaze data to produce "...discrete tokens ... that are claimed to approximate more closely the user's intentions in a higher level user-computer dialogue" (ibid., p. 166).

The "Little Prince" application (cf. section 4.3.1), includes an "interest module," that also computes the object looked at, and increments a "tally" bound to that object. This way, the user's gaze-pattern is used to produce aggregate data desribing not the user's gaze at one single point in time, but rather the user's interests in a time interval.

We believe that the ideal eye-gaze application should use several "dimensions" of the eye-tracking data. As described in section 3, the final eye-movements are caused by many different cognitive processes, so different aspects of the user's cognitive state are exposed by different features of the gaze pattern. As an example, the rate of mental activity can be detected by the rate of eye-movement (Kahneman 1973, p. 65), and Ponsoda et al. (1995) have analysed the saccade directions as opposed to the traditional fixation pattern analysis; they concluded that information on the proportion of diagonal saccades gives additional, non-superfluous information on the cognitive processes.

The "Midas Touch" Problem.

As described, the eyes cannot be used directly as a mouse, because the eyes are never "off." Thus one of the main problems when using the eye-gaze for selection purpose is to somehow combine it with a "clutch" that can engage/disengage eye-gaze control. A good clutch should be quick to operate, not increase the cognitive load unnecessarily and not disturb the user's gaze-pattern, because the user will often be looking at some object when she wants to engage eye-tracking, and it would thus slow down the communication if she had to move her eyes to do it.

One-way Zoom Problem.

As eye-gaze in many cases can be used to determine the object of interest, it is advantageous to use eye-gaze control for selecting an object (or a small group of objects) in a display for further examination by zooming in on it. If this can be repeated in any zoomed area, it is basically equivalent to hypertext with graphics, such as the World-Wide-Web. Using eye-gaze for "clicking" through hypertext is certainly a direct and efficient way of following links, but it lacks an important feature: the ability to go back to the previous level, i.e. zooming out.

There seems to be no natural way to implement a zoom-out function, but reasonable requirements for a zoom out function are again that it should be quick to operate, simple to operate and preferably be a natural thing to do when wanting to zoom out.

The Midas Touch Problem and the One-way Zoom Problem are hard to solve if one does not use the non-command approach, but use eye-gaze directly as a replacement for the mouse. Both could naturally be solved by the use of a manual button, but this seems like a clumsy solution and has shown itself to be inferior in practise, in relation to a latency-time solution (Jacob 1993). The latency-time solution is not ideal either, though; it is not quick to operate-especially for trained users-and it can increase the cognitive load on the user because she must be careful not to rest her eye too long on a selection object she does not wish to select.

A solution to the One-way Zoom Problem might be to have a discreet sign below the border of the screen, perhaps with an icon or a text saying "Zoom Out." Fixating this sign should then initiate the zooming out. Empirical studies must show whether this is a quick and simple solution-as for the naturalness, it is natural for the human eye-gaze to be lowered when wanting to go from focusing on something far away to something close-up on the ground.

Perhaps the two problems can be solved by using the gesture mode, not only the hands, but by using the whole body as a kind of joystick; moving one's head back more than a simple nod could initiate the zooming out-and leaning towards the screen could then zoom in (select the object), circumventing the Midas Touch Problem.