Humans optimally weight stereo and texture cues to estimate surface slant

Abstract

Purpose: The optimal, linear way to combine information from multiple cues is to weight cues according to their reliability. Previous work has shown that subjects adapt their cue weighting in response to global parameters that affect cue reliability (e.g. viewing distance for stereo). Cue reliability, however, can also vary with the value of the parameter(s) being estimated. Texture is a highly reliable cue for surface slant at high slants, but is much less reliable at low slants. We tested whether subjects' weighting of texture and stereo cues varies as a function of surface slant in an optimal manner. We further tested whether inter-subject differences in cue weights are predicted by the relative reliability with which subjects individually discriminate slant from the two cues in isolation. Method: In a first experiment, we used a 2-AFC task to measure thresholds for discriminating slant from texture alone (monocular views of random tiled textures), stereo alone (binocular views of random dot patterns) and texture and stereo combined (binocular views of random tiled textures). In the second part, we used a similar psychophysical procedure to measure points of subjective equality between stimuli with small cue conflicts and those without, as a way of measuring the weights given by subjects to stereo and texture cues. Results: Discrimination thresholds for slant from texture decreased relative to thresholds for slant from stereo as surface slant increased. As predicted by theory, subjects' texture weights correspondingly increased with increasing slant. Moreover, after subtracting the main effect of slant on texture and stereo weights, individual subjects' relative stereo / texture thresholds accounted for 60% of the remaining variability in cue weights. Conclusions: Subjects weight stereo and texture cues for slant in a manner consistent with the uncertainty of their of their own estimates of slant from the cues, as predicted by an optimal model.