Abstract
Purpose: We asked whether humans can prior probabilities used used to interpret a pictorial depth cue from visual information alone. In particular, we asked whether subjects learn to down-weight the compression cue (the bias to interpret elliptical figures as slanted circles) relative to stereoscopic cues in an environment containing randomly shaped ellipses. Methods: Subjects viewed stereoscopic images of elliptical figures and adjusted a 3D line probe to appear perpendicular to the figures. Test stimuli were near-circular ellipses at a slant of 35°. Non-test stimuli were ellipses at slants between 15° and 45°. Test ellipses were designed to suggest 5° conflicts between the compression cue and the stereoscopic cues. In experiment 1, non-test ellipses were all circles. In experiment 2, they had random aspect ratios. We used subjects' probe settings for test stimuli to compute cue weights in each of five successive daily sessions. Results: In experiment 1, subjects gave equal weights the compression cue (0.5) across all sessions. In experiment 2, subjects' compression cue weights shrank from 0.5 to 0.33 from the first to last session. Similar results were obtained when visuomotor performance (placing an object on the slanted stimulus) was used to estimate subjects' cue weights. Conclusions: Humans can adapt the prior distribution of object properties used to interpret a pictorial cue from visual information alone. These adaptations show up implicitly in the effective weight that subjects give to the cue. We describe a Bayesian learning model that accounts for subjects' performance.
Research supported by NIH grant EY-13319