Everyday perception
and
its restricted relevance to perception research

In daily life, we encounter dynamic scenes as well as static scenes. Both dynamic and static scenes may be visually ambiguous, but in daily life, this ambiguity may often be resolved by changes in view. In dynamic scenes, such changes in view may be caused by physical changes in the scene itself. For instance, if you encounter the scene in the next left-hand figure and if you are asked to assess how many windows the front of the building has, then you might answer "12 windows" (as, in some countries, children are taught to answer) even though, factually, you see only 8 whole windows plus a three window-like parts that might extend into many different things behind the truck. If the truck moves forward, however, the visual ambiguity is resolved, yielding a potentially unexpected answer.




Furthermore, notice that seeing organisms usually are able to move too, that is, the ability to see seems to have co-evolved with the ability to move. This implies that, also in case of static scenes, seeing organisms can often move around to obtain different views of a scene. The next example illustrates that, just as above, new views may lead to an update of the initial percept.

1. You take a first glance at a static scene		2. You probably interpret it like this
3. You move and take a second glance		4. You might see this, leading to an update of your initial percept

Formally, such a visual update process can be modeled straightforwardly by a recursive application of Bayes' Rule (see Occam, von Helmholtz, and Bayes):

• For a first view, one may multiply prior probabilities (quantifying viewpoint-independent aspects of candidate interpretations) and conditional probabilities (quantifying viewpoint-dependent aspects of candidate interpretations in relation to this first view) to obtain the interpretation with the highest posterior probability according to Bayes' Rule.
• For a second view, the just obtained posterior probabilities of candidate interpretations can be taken as their new prior probabilities which, together with the conditional probabilities for the expanded set of views, can be used to obtain new posterior probabilities by again applying Bayes' rule -- and so on for each additional view.

This recursive application of Bayes' rule implies that the effect of the initial, or first, priors gradually fades away because the priors are continuously updated on the basis of the conditionals which, thereby, become the decisive entities. Generally, just a few different views suffice to disambiguate the initial percept and to home in quickly on the "true" interpretation, that is, on the interpretation which, under the employed priors and conditionals, continues to get the highest posterior when the set of views is expanded further.


On the one hand, the foregoing is of course convenient to model everyday perception, be it by humans or by robots. On the other hand, research into human perception focuses primarily on properties of the human visual system, and one of the most interesting questions in this respect is the question of which priors it uses for a first view. In the dynamic situations above, these first priors may be less relevant but, of course in combination with first conditionals, they are relevant in completely static situations -- which are also part of daily life. Investigating the combination of first priors and first conditionals may reveal much about human vision (see also T-junctions).

For instance, in the recursive application of Bayes' rule above, it is left open where the employed probabilities come from. To merely simulate everyday perception, they might be based on empirical data. To explain it, they might reflect real probabilities of occurrence in the world, as considered by the Helmholtzian likelihood principle (though see Bertrand's paradox), or artificial probabilities derived from descriptive complexities, as considered by the Occamian simplicity principle. These two principles are probably far apart regarding the viewpoint-independent priors but seem close regarding the viewpoint-dependent conditionals (see Occam, von Helmholtz, and Bayes and Object versus viewer). Because the conditionals are the decisive ones in the recursive application of Bayes' rule, this implies that either principle would do in dynamic situations. This implies in turn, however, that such situations are not suited to investigate the question of which of these two principles guides the visual system and that, to this question, the first priors are most relevant.

Hence, in sum, modeling everyday perception in dynamic situations is of course a very interesting issue, but in human perception research, a more relevant issue is the underlying question of which first priors and conditionals are used by the visual system, and to investigate this question, static situations are better suited than dynamic situations.


For an extensive discussion on these issues, see Psychological Bulletin 2000
For a brief discussion on these issues, see In the Mind's Eye 2007
For an updated brief discussion on these issues, see Acta Psychologica 2011