Frames of reference for the light-from-above prior in visual search and shape judgements

Faced with highly complex and ambiguous visual input, human observers must rely on prior knowledge and assumptions to efficiently determine the structure of their surroundings. One of these assumptions is the ‘light-from-above’ prior. In the absence of explicit light-source information, the visual system assumes that the light-source is roughly overhead. A simple, low-cost strategy would place this ‘light-from-above’ prior in a retinal frame of reference. A more complex, but optimal strategy would be to assume that the light-source is gravitationally up, and compensate for observer orientation. Evidence to support one or other strategy from psychophysics and neurophysiology has been mixed. This paper pits the gravitational and retinal frames against each other in two different visual tasks that relate to the light-from-above prior. In the first task, observers had to report the presence or absence of a target where distractors and target were defined purely by shading. In the second task, observers made explicit shape judgements of similar stimuli. The orientation of the stimuli varied across trials and the observer’s head was fixed at 0, ±45 or ±60°. In both tasks the retinal frame of reference dominated. Visual search behaviour with shape-from-shading stimuli (SFS) was modulated purely by stimulus orientation relative to the retina. However, the gravitational frame of reference had a significant effect on shape judgements, with a 30% correction for observer orientation. In other words, shading information is processed quite differently depending on the demands of the current task. When a ‘quick and dirty’ representation is required to drive fast, efficient search, that is what the visual system provides. In contrast, when the task is to explicitly estimate shape, extra processing to compensate for head orientation precedes the perceptual judgment. These results are consistent with current neurophysiological data on SFS if we re-frame compensation for observer orientation as a cue-combination problem.