Abstract
Gradients such as shading and interreflections contain information about shape and scene structure. As with other cues to shape, the information in the gradients is ambiguous and the visual system requires constraining information to make use of it. We investigated whether sufficient information to disambiguate gradient depth cues is available as visual priors, or if aspects of the visual scene (e.g. light position), must be learnt in order to effectively use the gradient information. Stimuli consisted of rendered, physically accurate, cards with one side white, one red, separated by a vertical fold to form a concave or convex shape. Participants set shape estimates for stimuli at a range of card angles. Cues to depth other than gradients, such as binocular disparity and object outline were minimised by the use of a viewing aperture. Four observers received only a verbal explanation of the task (‘no training’), six saw a 30s video showing the full range of shaded stimulus angles with a congruent object outline and correct settings (‘full training’) and four saw a similar video but with incongruent gradients corresponding to the inverse shape (‘incongruent training’). Observers only participated in one condition. Results show that observers could not make accurate shape settings without seeing a representation of the stimulus beforehand. The ‘full training’ condition was however sufficient for observers to make consistent and unambiguous settings. This result could arise if observers made settings based on a mapping of gradients learnt from the video, but the results of the ‘incongruent training’ do not support this as three of the four participants made settings similar to those in the ‘full training’ condition. Our results indicate that that the visual system is capable of using complex illumination gradients as a depth cue, but that priors alone are not sufficient to disambiguate the gradients.
Meeting abstract presented at VSS 2012