A natural constraint is a key notion in the natural computation school’s approach to explaining visual processing (Marr, 1982; Richards, 1988; Ullman, 1979). According to this school, the key problem facing vision is underdetermination: the information contained in the proximal stimulus (e.g., retinal responses) is insufficient to uniquely specify or reconstruct the distal stimulus that produced the proximal stimulus. That is, stimulation in the eye is not sufficient to build a model of the world. How does one solve the problem of underdetermination? One approach would to be to provide information in addition to that present in the proximal stimulus. This information could be used to uniquely determine what distal stimulus led to the creation of the current proximal stimulus.
Natural constraints provide this additional information. They are properties of the visual world (they are not psychological properties) that are almost always true. They can be used to solve the problem of underdetermination by choosing the distal stimulus that 1) could have caused the proximal stimulus and 2) is consistent with a particular set of natural constraints. One general natural constraint is that a wide variety of visual properties (e.g., luminance, color, depth, motion) vary smoothly that is, a property measured at one point in the visual field leads to a very similar measurement at a nearby location. So, one could solve a problem of underdetermination by choosing the distal stimulus that maximized the smoothness of whatever properties were being measured by a visual process.