Abstract
Human vision is sensitive to both first-order information (luminance modulations, LM) and second-order information as conveyed by modulations of the local luminance contrast (CM) or amplitude (AM) of a visual texture. However, the role of second order vision is not yet clear. Johnson and Baker (JOSA-A, 21, 913–925, 2004) found that first- and second-order cues are correlated in natural images, but Schofield (Perception, 29, 1071–1086, 2000) showed that the sign of this relationship varies from image to image, suggesting that the phase relationship of the two cues may be an important environmental factor. We studied perceptual responses to mixtures of LM and AM (Schofield et al, Vision Research, 46, 3462–3482, 2006). Observers saw in-phase mixes (LM+AM) as strongly undulating surfaces via shape-from-shading. Negative (anti-phase) mixes (LM-AM) were also seen as undulations when presented alone. However, when LM-AM was presented together with LM+AM in a plaid the anti-phase cue was seen as flat - like strips of material laid across an undulating surface. Crucially, the perception of LM-AM depends on the context in which it is presented. It is tempting to suppose that these contextual effects have a high-level origin, but we now show that our results can be predicted by a simple, filter-based, bottom-up model. Images are passed through two linear channels with orthogonal preferred orientations and through two non-linear, filter-rectify-filter channels whose second-stage properties match those of the linear channels. Linear and non-linear channel outputs are combined by addition within each orientation and then subjected to mutual, cross-orientation inhibition prior to a final summation stage. The model outputs a shading map from which we estimate surface undulations. Both LM+AM and LM-AM produce undulations when presented alone, but for plaids, the LM+AM component dominates the LM-AM component and the latter cue does not produce undulations.
Supported by EPSRC grants GR/S07254/01 and GR/S07261/01.