The psychophysical data reported here bear on how the boundaries of an object's retinal image are processed in early vision. We propose that the visual system contains a mechanism sensitive to four relationships between two local stimuli some distance apart and that the output of this hypothetical mechanism encodes and labels orthogonally the four relationships. We measured the just-noticeable difference in the orientation difference between two test lines as well as the just-noticeable differences in their mean orientation, mean location, and separation. A pair of noise lines was located between the two test lines. By arranging that trial-to-trial variations in the orientation difference, mean orientation, mean location and separation of the test lines had zero correlation with each other and with trial-to-trial variations in the corresponding variables for the two noise lines we could demonstrate that psychophysical responses were based on the task-relevant variable and that, for each of the four task-relevant variables, all task-irrelevant variables were ignored. The finding that responses to the test lines were unaffected by the noise lines implies that discriminations were not influenced by first-stage spatial filters with strictly local receptive fields that responded to both test lines. Because these findings held for a presentation duration of 20 ms we can exclude the possibility that observers compared the two test lines by shifting either fixation or attention. We propose that, rather than by attending to two different locations, the test lines were selected by attending to the output of the long-distance comparator whose 'separation' label signaled the largest magnitude. The above proposals can account for several previously reported phenomena. More generally, an array of the proposed long-distance comparators constitutes a system that may be capable of specifying the shape, size, location and implicit orientation of an object's retinal image.