Vision Modelling. Part Six. Modelling Eye Tracking
1 Excursion into the eye — 2 Perception — 3 Geometry of vision — 4 Eye tracking — 5 How to catch the gaze — 6 Modelling eye tracking
The sixth and final part. An account directly of the principles of modelling eye tracking, and also of how this may be applicable to real life. In this part, at last, the figures and principles set out in the third, fourth and fifth posts will come in useful.
Which models of eye tracking are of interest?
As was told in the fourth part, there is a direct connection between the time spent viewing a region of an image and the degree of attention devoted to that same region by a person’s consciousness. At the very least, if a person does not even realise what they were looking at, they will certainly remember it!
Therefore one kind of modelling is the creation of a heat map of viewing duration, or information about what potential for prolonged study a given region of an image possesses.
Another interesting kind of model is the dynamics of saccades: a diagram showing toward which point a person’s gaze will strive if at present it is looking at the current one. By analysing this kind of diagram one can understand which parts of the image seem the most important during viewing. And by combining the diagram of dynamics with the heat map one can understand which side details are capable of distracting a person’s gaze from what they consider most important.
The bounds of applicability of modelling
Only psychologically neutral images lend themselves well to the modelling of saccades: emotionally coloured pictures shift the centres that draw attention considerably. For example, when looking at a crying child a person’s gaze will inevitably devote more attention to studying the expression of the face than to incidental contours.
This same factor means that a person’s gaze is entirely individual: if a particular person is not indifferent to pigeons, then when looking at a photograph with pigeons they will undoubtedly move their gaze differently from a person indifferent to birds.
The presence in a picture (even a psychologically neutral one) of small but recognisable contours (people’s faces, symbols of text, and so on) sets modelling the task of a preliminary recognition of the images of those contours. In the program that served as the occasion for writing this manual there is no recognition at all, only an estimate of the complexity of perceiving a contour. But the conclusion that the quality of image recognition substantially improves the correspondence of the model to data obtained from real eye tracking, while its absence may lead to fatal mismatches, must necessarily be drawn.
Modelling works well for situations of surveying a picture without any purpose, without a task having been given. When an image is viewed with a definite purpose, the role of image recognition increases, affecting the strength of perception of contours.
Where can this be applicable?
Upon becoming acquainted with the bounds of applicability of the models, a logical question arises: “And where, with such limitations, can this be applicable in real life?”
And it is applicable in a wide class of tasks, above all those connected with the analysis of text. As is well known, the perception of text is strongly influenced by its formatting, the size of the empty spaces between paragraphs, words and characters, the design of the typefaces, and so on.
Text, initially, before processing, is seen as a psychologically neutral picture ideally suited to analysis by the methods of modelling saccades. That same heat map shows which words, paragraphs and other textual units are most inclined to be remembered, because, owing to the formatting of the text, a greater amount of attention is paid to them. And on the other hand one can understand which blocks of text, because of that same formatting, fall out of attention as a consequence of landing in a zone of indistinguishability.
Besides the analysis of the perception of texts, the models can work quite well for finding optimal regions of perception (in order to show something — an advertising block) or of non-perception (in order to hide something — a contact telephone number) ;)
And one should not altogether exclude the class of tasks of the rapid analysis of images for the purpose of usability testing. Access to a “grown-up” eye-tracking system and a dozen people from a focus group is often an impossible luxury.
How it works
The modelling of saccade vectors at the simplest level (that is, the way it works in my case) consists in solving two problems:
- The computation of a certain “strength of perception” for a specified area.
- The finding and extraction of gaze traps on the basis of information about the “weight” of a centre of attention.
The classical solution of the first problem is carried out by the following sequence of actions:
- Extraction of the required area, in accordance with the angle of the eye’s optical perception;
- Normalisation of the extracted area in accordance with the parameters of a person’s visual acuity (which is why a Gaussian blur is not entirely applicable here);
- Extraction of contours using frequency analysis and perceptual thresholds (at this stage what matters to us is not so much the contours themselves as their approximate outlines; Gabor filters fit rather well);
- Detection of breaks in the continuity of contours and normalisation of the points of discontinuity by Glass’s algorithm (in order to find the vectors of subjective contours and to supplement the objective contours with them);
- Computation of the weight of the zone on the basis of the number of changes of direction, the scatter of the directions of the contour vectors, and the number of break points.
In fact, since the weight of a zone is merely an estimate of the complexity of the contours that lie in it, and the complexity of contours is a certain composition of the scatter of the directions of the vectors and of the break points. And for the second problem we do not need exact numbers, but their relative weight. Therefore instead of a function of the complexity of a contour one may take its derivative. And the derivative, for its part, is proportional to a composition of the coordinates of the ends of segments in radial coordinates. Accordingly, the task of determining the complexity of a contour can be reduced to finding the number of intersections of the contours with a large number of radial chords.
Experience shows, however, that at small values of the radius (in the case of 96 dpi the radius of the zone of sharp vision will be about 18 pixels, and the region under study, taking peripheral vision into account, 57), 9 radial chords with a step of 20 degrees are sufficient for an acceptable computation of the weight. Decreasing the angular step raises the precision of determining the order of the weight, but only slightly compared with the results obtained by the classical model of computing the weight.
The same experience shows that for different dpi and, accordingly, different radii of the zones of vision, it makes sense to use the same 18-ray model, with a preliminary normalisation of the region in the form of reducing it to a circle with a radius of 30 pixels.
Why does greater precision of computation in any case fail to lead to a better result? Because the very operation of extracting contours possesses a certain error, and the computational error of the 18-ray model is comparably small relative to the error of the method itself.
The second problem — the problem of extracting traps for the gaze — may be solved in the following way:
- Scanning the entire area of the picture and extracting the zones with maximal weights
- Building a graph from the points found and assigning weights to the links between them
- Solving a standard problem on the graph
1 Excursion into the eye — 2 Perception — 3 Geometry of vision — 4 Eye tracking — 5 How to catch the gaze — 6 Modelling eye tracking