We perceive the visual world as made of objects of different shapes, sizes and colors. Some may be smooth, shiny and reflective, whereas others are rough and uneven; some may be in shadow, while others are brightly lit. The brain must identify and distinguish all of these different features to build an accurate, three-dimensional model of the environment.

Information about any visual feature originates as light bouncing off an object and entering the eye, which then captures the reflected light and focuses it onto the retina. There, cells generate electrical signals for the brain to process. However, different types of visual features can result in the same pattern of activity. The brain must rely on prior knowledge and educated guesses to disentangle the contributions made by different features, but we know little about the processes that make this possible.

Here, Mooney et al. examine how the visual system can tell whether an object is blurry, or if it presents the smooth light-to-dark shading that can accompany curved shapes. The experiments show that images of shaded curved surfaces can appear blurry even when they are fully in focus. However, adding a specific type of sharp edge, called a bounding contour, eliminates this illusion. This suggests that the brain uses these sharp edges to judge whether an image is in focus. In fact, adding bounding contours can trick the visual system into perceiving a blurry image as sharp.

Understanding how the human visual system interprets images could lead to advances in computer vision. Artificial vision systems – such as those used in face or license plate recognition – must determine which parts of an image are in focus before attempting to extract visual information. Identifying the cues that enable the human visual system to solve this problem could help to train computers to do the same.

An image of a surface is a product of its three-dimensional (3D) shape, the reflectance and transmittance behavior of its material, the surrounding light sources (the ‘light field’), and the focal parameters of the imaging lens. All of these sources are conflated in the light that reaches the eyes, yet we nevertheless perceive distinct impressions of shape, material, illumination, and focus. One of the fundamental goals of mid-level vision research is to understand how the visual system extracts these different sources of structure.

Most research into this problem has focused on how the visual system extracts environmental sources of structure – 3D shape, reflectance (color, lightness, gloss), and surface opacity (transparency, translucency, and subsurface scattering). But the focal properties of single-chambered eyes also contribute to the optical structure projected to the retinae. Eyes that utilize refraction to generate a focused image are subject to depth of field defocus, which causes some image regions to be blurred. Optical defects can also cause images to be globally blurred for the increasing number of people that require optical correction. Although there has been a significant body of research into cues that affect the severity of perceived blur (Ciuffreda et al., 2006; Crete et al., 2007; Ferzli and Karam, 2006; Pentland, 1987; Tadmor and Tolhurst, 1994; Webster et al., 2002) and the role of depth of field defocus as a cue to depth (Held et al., 2010; Marshall et al., 1996; Mather, 1996; Mather, 1997; Mather and Smith, 2000; Mather and Smith, 2002; O'Shea et al., 1997; Watt et al., 2005), it is still unknown how the visual system distinguishes blur from environmental sources of low spatial frequency image structure.

The computational problem of discriminating optical defocus from environmental sources of low frequency structure does not appear to have been explicitly addressed previously. This may be due to the absence of empirical evidence that the visual system can misattribute image gradients produced by environmental sources to defocus or misattribute gradients produced by defocus to environmental sources. Here, we provide evidence of both. We show that defocus can be experienced in fully focused images and that optical defocus can be misperceived as distortions in the perceived 3D shape of smoothly shaded surfaces.

Consider the surface depicted in Figure 1, which was created by illuminating a smooth (i.e., differentiable) Lambertian (‘matte’) surface with a collimated light source. The surface has shallow surface relief to avoid the formation of sharp attached shadows and is viewed along the axis of relief to avoid the formation of self-occluding contours. Although rendered as a fully focused surface, this image elicits a strong perception of blur; while some 3D shape from shading may be perceived, the surface appears ‘contaminated’ by optical defocus.

The perceptual conflation of low frequency shading gradients and optical focus does not appear to have been previously reported. Most research into the perception of shading has attempted to understand how shading provides information about 3D shape using images where it was assumed or somehow ‘known’ that the intensity gradients were caused by focused patterns of shading. The potential conflation of shading and blur may have been overlooked because of the particular surface geometries and viewing conditions that were used in these studies. Most previous work on shape from shading has studied images that contained sharp contours generated by either smooth self-occlusions or abrupt bounding contours (e.g. Horn and Brooks, 1989; Koenderink et al., 2001; Mingolla and Todd, 1986; Pentland, 1984; Ramachandran, 1988; Todd et al., 1996), while experiments that have used ‘terrain’ surfaces similar to Figure 1 have predominantly been investigations of illumination perception (Koenderink et al., 2004; Koenderink et al., 2007). This suggests that the visual system may exploit sharp contours in generating percepts of focus when viewing low frequency shading gradients. If so, it should be possible to eliminate the perception of illusory blur in Figure 1 by introducing sharp bounding contours.

Figure 1

Download asset Open asset

We informally tested this hypothesis by constructing the images depicted in Figure 2. The image on the left (2A) was generated by intersecting the surface in Figure 1 with a gray plane parallel to the direction of its relief and occluding all surface regions beyond the depth of that plane (a particular form of ‘planar cut’ dubbed a ‘level cut’). Informal observation suggests that this manipulation enhances the perceived 3D shape of the surface and completely eliminates the perception of blur. Note also that the contour appears to be unambiguously ‘owned’ by (attached to) the shading gradients rather than the homogeneous gray regions. Now consider the image in Figure 2B, which was created by treating the gray regions in Figure 2A as parts of a single gray ‘mask’ and rotating it 180 degrees over the shaded surface. This image contains the exact same contours as Figure 2A, but the gray regions occlude different portions of the shaded surface. There are a number of striking perceptual differences evoked by comparison of Figure 2B and Figure 2A. First, whereas the perceived shape from shading is enhanced in Figure 2A, the perception of shape from shading is impaired by the contours in Figure 2B. A second difference is that the border ownership of the contour in Figure 2B is ambiguous; whereas the contours in Figure 2A appear unambiguously attached to the shading gradients, the contours in Figure 2B can appear attached to either side. The way the border ownership is perceived can have a dramatic impact on how the gradients are perceived. When the contours appear attached to the shading gradients, the gradients are perceived as unstructured 2D ‘noise’ (such as variations in pigment) without any clear sense of 3D shape or optical focus. But when the contours appear attached to the gray mask, the shaded surface appears as a partially occluded shaded surface that is just as blurred as the surface in Figure 1.

Figure 2 with 2 supplements see all

Download asset Open asset

What is responsible for the striking perceptual differences observed in Figure 2A and Figure 2B? Why does one set of contours eliminate the perceived blur, enhance perceived 3D shape, and unambiguously determine the side of the contour that is figure, while the other does not? We suggest here that the perceived focus and enhanced perception of 3D shape arises because the level cut images approximate the geometric and photometric image properties generated by smooth self-occluding contours. More specifically, we will argue that there are two constraints that play a causal role in eliciting the perception of focus in images of shaded surfaces: photogeometric constraints that arise generically along smooth self-occlusions; and the attachment of shading gradients to convex surfaces. We consider each in turn.

The first constraint arises from the physics of surface reflectance and the projective geometry of smooth self-occluding rims. Both shading intensity and bounding contour shape depend on the same environmental property – the local 3D shape of the surface – and are therefore inherently linked. This link makes it possible to combine two well-known constraints about the shape of the contours generated by smooth self-occlusions and the shading intensity into a novel constraint (Marlow et al., 2019). First, it is known that the 3D shape along a smooth self-occluding rim can be derived from its 2D image contour: the local slant of the contour relative to the observer (i.e., how much it deviates from fronto-parallel) is constant (approximately 90 degrees), and the tilt of the surface (the direction in which it slants away from the observer) is specified by the orientation of the rim’s image contour (Barrow and Tenenbaum, 1978). Second, it is known that local 3D surface orientation is primarily responsible for shading intensity (Horn and Brooks, 1989), subject to some modification by vignetting and/or interreflections (Langer and Zucker, 1994). The mutual dependence of contour orientation and shading intensity on the same local 3D surface orientation causes shading to covary with the orientation of the rim’s projected image contour: for a Lambertian surface illuminated by a collimated light source, intensity will decline as a cosine function of contour orientation relative to the brightest point along the contour (i.e., the orientation most closely aligned with the illumination direction). This relationship holds exactly for surfaces illuminated by collimated light sources (see Materials and methods), and we have previously shown that it is statistically robust in natural light fields that contain multiple sources of illumination (Marlow et al., 2019). We refer to this relationship as the orientation-intensity covariation of the contour and its adjacent shading.

The covariation of contour orientation and intensity along smooth self-occluding rims generalizes to other types of bounding contours that have been shown to affect perceived shape from shading, such as ‘planar cuts’ (i.e., contours formed by ‘slicing’ a shaded surface with a plane; Marlow et al., 2019). The level cut image in Figure 2A is one example of a planar cut. The orientation-intensity covariation that arises along planar cuts of shaded surfaces is similar to that generated by smooth self-occluding rims (see Materials and methods). The relationship between contour orientation and shading intensity along the contours in Figure 2 is depicted in the plot below each image. Note that Figure 2A exhibits a clear cosine-like relationship, whereas Figure 2B exhibits no covariation at all. However, this covariation is generally weaker for planar cuts than for self-occluding rims: more than one 3D surface orientation can generate the same 2D contour orientation, which means that identically oriented planar cut segments can project different shading intensities in the image (as can be seen upon close examination of Figure 2A). Nonetheless, we previously showed that planar cuts exhibit a robust orientation-intensity covariation (apart from a few degenerate cases; Marlow et al., 2019), which suggests that this covariation could provide a reliable cue that the visual system uses to identify the bounding contours of shaded surfaces.

The second property that links the level cut image in Figure 2A to images containing smooth self-occlusions is that they are both globally convex (Koenderink, 1984). If the covariation of intensity and contour orientation along sharp bounding contours is sufficient to explain the perception of focus in shaded surfaces, then it should not matter whether these contours bound a convex or concave surface. The importance of convexity can be assessed by constructing a level cut surface that removes all of the convex surface regions that appear in front of the cut and displaying only the ‘valleys’ (Figure 2C). This stimulus exhibits the same orientation-intensity covariation, but is inherently ambiguous: it can be perceived as a convex surface illuminated from below or as a concave surface illuminated from above (e.g. Ramachandran, 1988; see Liu and Todd, 2004). There are two well-established biases that determine how such ambiguities are resolved: a bias to perceive the illumination as coming from above (Belhumeur et al., 1999; Koenderink et al., 2001) and a bias to perceive surfaces as convex (Hill and Bruce, 1994; Langer and Bülthoff, 2001). These two biases are aligned in Figure 2a, which is presumably why this surface is perceived as a stable convex surface illuminated from above. However, these biases are in conflict in Figure 2C, causing some perceptual bistability. Informal observations suggest that perceived focus depends on the perceived convexity or concavity of the surface. When the surface appears convex, no clear percept of blur is experienced; but when it appears concave, the surface appears blurred. This implies that the strength of the orientation-intensity covariation along the contours is not the sole determinant of the perception of shading, focus, and contour attachment; the convexity of the surface is also critical.

Our previous work showed that photogeometric cues along smooth self-occlusions and planar cuts of 1D luminance profiles provide information that predicts when identical luminance gradients are perceived as surface shading of 3D surfaces. The experiments described below were designed to psychophysically assess the relationship between bounding contour orientation, shading intensity, and contour sharpness on the perception of surface shading, optical defocus, and border ownership.

The goal of Experiment 1 was to test if sharp image contours cause bounded shading gradients to appear more focused when contour orientation covaries with the intensity of the bounded gradients. This covariation is typically exhibited along contours generated by self-occlusions. Self-occlusions feature prominently in prior work on shape from shading, but their contours are difficult to manipulate without altering the geometry of the entire shaded surface (and hence the shading gradients). However, it is possible to approximate the photogeometric behavior along smooth self-occlusions with planar cuts. This was accomplished by slicing the bumpy plane depicted in Figure 1 with a plane oriented perpendicular to the axis of surface relief (a ‘level cut’; see Materials and methods for details). The resulting level cut contours exhibit an orientation-intensity covariation similar to the covariation along self-occluding contours, but unlike self-occlusions, level cuts can be generated anywhere on the surface and place no constraints on local surface curvature. The photogeometric constraints of self-occlusions and level cuts are described in further detail in the Materials and methods.

To assess the role of the photogeometric covariation along bounding contours in the perception of 3D shape, border ownership, and optical focus, three sets of level cut contours were created by intersecting the surface in Figure 1 with a fronto-parallel plane at different depths along the axis of surface relief. Each set of contours was used to generate four homogenous gray masks. In the first condition, each mask occluded all surface regions that lay at a greater depth than the level cut, leaving only the shaded peaks visible (e.g. Figure 2A). The visual system’s biases toward interpreting shaded surfaces as convex and top-lit are aligned in this image, and the visible surface regions consequently appear unambiguously convex; we therefore refer to this condition as the ‘convex’ level cut condition. In a second condition, the same contours were used to generate a complementary mask that removed all regions in front of the level cut, leaving only the shaded valleys visible (e.g. Figure 2C). The visual system’s convexity and illumination biases conflict in this image; the shaded regions may appear as concave dents illuminated from above or convex bumps illuminated from below. This conflict can result in some bistability in the perceived illumination direction and convexity/concavity, so we refer to this as the ‘bistable’ level cut condition. The masks in the two ‘rotated’ conditions were created by rotating the level cut masks in the convex and bistable conditions (respectively) by 180° over the underlying gradients, which breaks the orientation-intensity covariation that occurs along the level cuts (e.g. Figure 2B). The depth of the intersecting plane used to define the contours determined the relative proportion of visible gradients in each masked image; the three depth values were chosen to create masks that preserved 25%, 50%, and 75% of the gradients. Note that the masks in the ‘convex’ conditions with 25% gradients visible are the complements of the masks in the ‘bistable’ conditions with 75% gradients visible, and vice versa. The luminance of each gray mask was set to the average of the original gradient image. The full set of twelve stimuli can be seen in Figure 2—figure supplement 1 and Figure 2—figure supplement 2.

If the visual system uses the orientation-intensity covariation between contours and shading to distinguish attached bounding contours from arbitrary edges, then the level cut contour conditions should induce stronger percepts of focus than the rotated contours, which have no relationship with the shading gradients. However, if the visual system is better at inferring contour attachment when the adjacent surface appears convex, then the covariation alone (which provides no curvature cues) may not fully eliminate the perception of blur. If this is true, then the contours in the ‘bistable’ level cut condition should produce weaker percepts of focus than the ‘convex’ level cut condition, as the bistable stimuli may sometimes appear concave.

Observers judged perceived focus in a paired comparison task and judged perceived surface curvature (convex, concave, or neither) in a separate three-alternative classification task (N = 20). The results confirm our hypotheses and informal observations of the stimuli (Figure 3) and were analyzed with an ANOVA and appropriate two-sided contrasts. Observers perceived the level cut stimuli (solid lines in Figure 3A) as appearing more focused than the rotated stimuli (dotted lines in Figure 3A), ${\displaystyle F\left(1,19\right)=64.68,p<0.001}$, 95% CI $[25.33,43.15]$, Cohen’s $d=1.85$, and further perceived the ‘convex’ level cut stimuli as more focused than the ‘bistable’ level cut stimuli, ${\displaystyle t\left(19\right)=7.05,p<0.001}$, 95% CI $[20.45,37.73]$, Cohen’s $d=1.62$. All observers perceived the ‘convex’ level cut surfaces as convex bumps on all trials (first plot in Figure 3B), whereas the gradients occluded by rotated masks were most likely to be perceived as neither bumps nor dents (third and fourth plots in Figure 3B). Perceived focus decreased as a function of increasing gradient visibility in the ‘bistable’ level cut condition, ${\displaystyle t\left(19\right)=7.07,p<0.001}$, and the rotated conditions, ${\displaystyle t\left(19\right)=5.05,p<0.001}$, but not the convex level cut condition, $t\left(19\right)=1.73,p=0.100$, which suggests that the convex level cut masks were the only stimuli that were perceived as (equally) fully focused. In the other three mask conditions, it is likely that observers simply preferred to select images in which less of the perceptually blurred gradients were visible.

Figure 3

Download asset Open asset

The data suggest that the effects of the different contours on perceived focus are mediated by the perception of convexity, which is strongest when the contours exhibit an orientation-intensity covariation with the shading and the convex interpretation is consistent with percepts of top-down illumination. Notably, perceived focus and perceived convexity simultaneously decreased in the ‘bistable’ conditions – but not the ‘convex’ conditions – as the ratio of visible gradients increased. There was a significant correlation between perceived focus and the proportion of observers who rated each stimulus as appearing convex in shape, ${\displaystyle \rho =0.904,p<.001}$. This correlation remained significant even when the unambiguous ‘convex’ mask conditions were excluded from the analysis, $\rho =0.777,p=0.014$. This finding supports our informal observation from Figure 2: contours generate stronger cues to contour attachment when they bound surfaces regions that appear convex, and these contours are therefore more likely to propagate focus cues (produced by their sharpness) to the shaded surface. This may also explain why prior studies using globally convex shaded 3D shapes have not observed any illusory gradient blur: the self-occluding contours of these stimuli not only exhibit a strong orientation-intensity covariation (Marlow et al., 2019) but also necessarily bound surface regions that have convex curvature in at least one direction. The bounding contours of surface regions that appear concave, however, can appear as occluding edges of the gray mask, similar to the percept that arises from the rotated contours (Figure 2B) when the gray regions appear to ‘own’ the contours and the surface appears occluded.

The results of Experiment 1 suggest that sharp contours that exhibit an orientation-intensity covariation with nearby shading (such as level cuts) provide information about image focus and enhances percepts of 3D shape within the shaded surface. In Experiment 2, we tested the importance of this covariation directly by parametrically varying the strength of the orientation-intensity covariation along contours. This was accomplished with images of smoothly shaded ‘ribbons’ on a gray background. These images were created by displaying only the shading gradients immediately adjacent to the mask contours used in Experiment 1. Thus, the only source of information about 3D shape is the relationship between the intensity and orientation along the ribbon. Sixteen unique ribbon paths (two pixels wide) were created by generating more smoothly deformed 3D surfaces and intersecting them with planes to produce level cut contours, but the ribbon was manually shaded in MATLAB according to the orientation of its path. This allowed us to create ribbons that exhibited perfect orientation-intensity covariation that could be parametrically decreased to any arbitrary value. Ribbons were constructed such that intensity decreased in relative intensity from 0.8 to 0.2 as a linear function of its orientation relative to 90° (the brightest point). The covariation was then progressively weakened by gradually adding increasing amounts of random low-frequency noise to the ribbon gradients. The gray background had a relative intensity of exactly 0.5.

Figure 4 depicts three example stimuli. The ribbons in the left panel exhibit a perfect linear orientation-intensity covariation, which induces a vivid percept of surface relief: some of the gray regions are perceived as ‘plateaus’ raised above the adjacent recessed regions. The Pearson correlation coefficient between shading intensity and contour orientation (relative to 90°) measured directly from each image is shown in the lower-right corner. In the center panel, the intensity of a different perfectly covarying ribbon has been mixed in equal proportion with low-frequency noise. A moderate degree of covariation remains, but it is not consistent across the image, and the overall impression of stepped relief is substantially weaker. In the right panel, ribbon intensity has been generated entirely by noise; no covariation is present, and no impression of 3D relief is apparent.

Figure 4

Download asset Open asset

Observers (N = 15) were shown randomly-generated ribbon images with thirteen values of noise proportion ranging from 0% (as in the left panel of Figure 4) to 100% (as in the right panel of Figure 4). These values were not evenly spaced but were instead selected to produce an approximately uniform distribution of correlation coefficients between 1 and 0 when the covariation was measured in each image. Each observer viewed all possible pairs of these thirteen noise values and were instructed to select the image that appeared more three-dimensional in each pair.

The results confirm our informal experience of Figure 4: 3D shape percepts monotonically decreased in strength as a function of increasing ribbon noise (left plot in Figure 5). This relationship was verified with a linear regression, with the likelihood of being selected as appearing more three-dimensional decreasing by approximately one percentile for every percentile increase in ribbon noise, ${\displaystyle b=-1.025,R^2=0.923,p<0.001}$. Further analysis revealed that this effect was mediated by the amount of orientation-intensity covariation present in the images: mean perceived 3D shape strength also decreased monotonically with the value of the Pearson correlation coefficient when the correlations measured from all presented stimuli were sorted into bins of width 0.1 (right plot in Figure 5). These findings do not directly address the issue of perceived defocus, but do support our hypothesis that the photo-geometric behavior occurring at the very edge of covarying contours (such as the level cut contours in Figure 2A) is sufficient to generate the vivid impressions of contour attachment observed in the ‘convex’ conditions of Experiment 1. The data also reinforce our previous findings on the importance of this covariation in generating percepts of 3D shaded shape (Marlow et al., 2019).

Figure 5

Download asset Open asset

The results of Experiments 1 and 2 suggest that the absence of an orientation-intensity covariation along contours can lead to the misperception of optical defocus and impair the perception of 3D shape. In Experiment 3, we investigated whether image structure that has been optically blurred by defocus can be misperceived as focused if there are sharp contours present nearby that exhibit an orientation-intensity covariation; that is we tested whether the presence of such contours can effectively mask the visibility of optically induced blur. We tested this by constructing two variants of the shaded terrain in Figure 1. One variant was similar to the level cut image, which contained sharp bounding contours a given relief height (middle of the bottom row of Figure 6), and therefore referred to as the level cut image. The other image was constructed by geometrically smoothing the sharp level cut contours of this image, resulting in the image depicted in the middle of the top row of Figure 6. As in Figure 1, the absence of sharp, intensity-correlated contours causes this surface to appear blurred even though it was rendered in full focus. We refer to this surface as the ‘smoothed’ condition. The difference in perceived blur experienced with these two images reinforce the importance of contour sharpness for inducing percepts of focused gradients: the absence of sharp edges in the smoothed condition greatly reduces perceived focus, even though the changes in surface curvature at the edges of the bumps are relatively abrupt.

Figure 6

Download asset Open asset

To assess sensitivity to optical focus, both shaded surfaces were subject to different degrees of optical defocus, with the focal length set to either the background (so the contours surrounding the bumps in the level cut condition remained sharp) or the tips of the bumps (so the contours in the level cut condition were blurred by defocus). The effects of increasing background blur and increasing foreground blur can be seen to the left and right of the center column in Figure 6, respectively. The effects of foreground blur on the level cut surface (bottom-right panels) are particularly striking: when the peaks of the surface are affected by defocus blur, the resulting changes in the image gradients do not appear to significantly change the perceived focus of the surface. In the absence of the level cut contours, however, the same defocus manipulation appears to increase apparent blur (top-right panels).

These informal observations were bolstered by psychophysical experiments that measured perceived focus for all ten stimuli using a paired comparison task (N = 10). The results (depicted in Figure 7) align with our informal observations of the stimuli in Figure 6 and were analyzed with appropriate contrasts within an ANOVA (the main effects of which are not relevant here). The vertical axis in Figure 7 represents the percentage of trials in which each stimulus was selected as appearing more focused, and the horizontal axis represents the different blur conditions. The blue and red lines depict perceived focus for the smoothed and level cut conditions, respectively. The data reveal that defocus blur can be misperceived as shading: the three stimuli with sharp level cut contours (i.e.the level cut condition with no blur or foreground blur) were perceived as the most focused, even when the peaks of the surface were actually blurred by moderate or severe foreground defocus. Statistical analysis revealed that foreground blur had a significantly larger (more negative) effect on perceived focus than background blur for the level cut surface, ${\displaystyle t\left(9\right)=20.74,p<0.001}$, 95% CI $[45.29,56.38]$, Cohen’s $d=6.91$, but not the smoothed surface, $t\left(9\right)=0.31,p=0.763$, 95% CI $[-11.49,8.71]$. Furthermore, the fully-focused smoothed surface with no contours did not significantly differ in perceived focus to the level cut stimulus with weak background blur, $t\left(9\right)=0.71,p=0.494$, 95% CI $[-9.28,4.84]$. These results demonstrate that the presence or absence of sharp contours attached to shading gradients can induce both misperceptions of optical focus or defocus (respectively) when they exhibit a systematic intensity-orientation covariation.

Figure 7

Download asset Open asset

The results of the preceding experiments suggest that the most important mediating factor in the perception of focused shading gradients is the presence of sharp bounding contours that exhibit covariation between their orientation and adjacent shading intensity. In Experiment 3, we found that optical blur was overestimated when these sharp, covarying contours were absent from the image. Actual changes in image focus, however, were effectively undetected by observers when the sharpness of nearby covarying contours was preserved. In Experiment 4, we tested whether this underestimation of optical blur occurs because the smooth gradient structure is misattributed to environmental sources: that is whether the optical effects of defocus are attributed to the 3D shape and reflectance properties of the defocused surface. Surfaces with higher microscopic roughness (e.g. glossy plastic or matte materials) will scatter incoming light in more directions and produce smoother image gradients instead of the sharp, detailed specular reflections produced by low-roughness surfaces (e.g. mirrors). This effect of surface roughness on image structure is further mediated by curvature: high-curvature surfaces vary in surface orientation more rapidly than low-curvature surfaces, and the shading or reflections they produce are therefore more compressed (i.e., have higher spatial frequency) in the image. We have previously demonstrated that the visual system can misperceive low-curvature, low-roughness surfaces as high-curvature, high-roughness surfaces (Mooney and Anderson, 2014), as both combinations of surface properties generate similar gradient structure. The local effects of optical blur on image gradients are similar to the effects of increasing surface roughness or decreasing surface curvature: all three of these physical transformations typically reduce the sharpness of image gradients exhibited by the surface. It is therefore likely that the visual system will have difficulty distinguishing these optical and environmental influences on gradient appearance when they do not produce simultaneous changes in contour sharpness or covariation strength.

We tested this hypothesis by creating three identically shaped surfaces with different reflectance properties: ‘matte’, ‘rough gloss’, and ‘smooth gloss’. All three materials have an identical diffuse shading component, but the two gloss conditions also contain a specular reflectance component. The amount of scattering in this specular component is greater in the ‘rough gloss’ condition, which consequently produces reflections with less detail than the ‘smooth gloss’ condition. The surface’s 3D shape had higher curvature than the shape used in Experiment 3 to increase the likelihood of generating measurable misperceptions of 3D shape. Level cut contours were created by intersecting the surface with a plane, as in previous experiments, but the matte gray planar surface itself was here included as part of the rendered scene rather than a mask added to the image afterward. The combined surface was rendered in a natural light field with cast shadows, inter-reflections, and chromatic information. This was done to test whether misperceptions of optical blur occur in more realistic viewing conditions, which are particularly important for the appearance of glossy materials (Fleming et al., 2003; Pellacini et al., 2000).

Each material condition was rendered with five optical defocus conditions, identical to the conditions in Experiment 3: the ‘foreground blur’ conditions blurred the peaks of the surface but preserved the sharp contours, and the ‘background blur’ conditions blurred the sharp contours, leaving the peaks unaffected. The fifteen stimuli are depicted in Figure 8, which has a similar layout to Figure 6. Each row depicts a different material and each column depicts a different focus condition. As background blur increases to the left of the center column, perceived focus appears to decrease for all three materials. As foreground blur increases to the right, the changes in gradient appearance are misattributed to transformations in material (the surface appears more matte) and 3D shape (the surface appears less curved).

Figure 8

Download asset Open asset

We measured perceived image focus, perceived surface gloss, and perceived 3D shape in three distinct tasks. Perceived focus and perceived surface gloss were each measured from the same group of observers (N = 10) in two separate paired comparison tasks. Perceived 3D shape was measured from expert observers (N = 5) for six of the stimuli (outlined in red in Figure 8) using a line of twenty ‘gauge figure’ probes across the prominent ridge on the left side of each image (red dots in central stimulus of Figure 8). These probe settings were integrated to form cross-sectional profiles of perceived relief (see Koenderink et al., 1992; Koenderink et al., 2001). The results accord with our findings in Experiment three and support our informal observations of the stimuli, which are described in separate sections below for each of the three measured properties.

Observer’s reports of perceived focus are depicted in Figure 9 and were analyzed with appropriate ANOVA contrasts, as in Experiment 3. The vertical axis is the percentage of trials in which each image was selected as appearing more focused. The horizontal axis plots the different blur conditions and each colored line represents a different surface reflectance type. The data indicate that background blur (moving from the central ‘no blur’ condition to the left) significantly reduced perceived focus for all three materials, ${\displaystyle t\left(9\right)=-39.62,p<0.001}$, 95% CI $[-71.86,-64.10]$, Cohen’s $d=13.21$. Foreground blur (moving from the ‘no blur’ condition to the right) also reduced perceived focus, ${\displaystyle t\left(9\right)=-9.75,p<0.001}$, 95% CI $[-29.92,-18.65]$, Cohen’s $d=3.25$, but to a significantly lesser extent than background blur, ${\displaystyle t\left(9\right)=-34.80,p<0.001}$, 95% CI $[-46.53,-40.85]$, Cohen’s $d=11.60$. The data also reveal interactions between blur type and material: in the background blur conditions, the surface with smooth gloss was perceived as significantly more focused than the surface with rough gloss, ${\displaystyle t\left(9\right)=3.58,p<0.005}$, 95% CI $[3.16,13.98]$, Cohen’s $d=1.19$, which was in turn perceived as more focused than the matte surface, $t\left(9\right)=4.27,p=0.002$, 95% CI $[4.03,13.11]$, Cohen’s $d=1.42$, but there were no significant differences in perceived focus between materials in the foreground blur conditions, $t\left(9\right)=-0.13,p=0.901$, 95% CI $[-6.67,5.95]$ (smooth vs. rough gloss) and $t\left(9\right)=-1.59,p=0.146$, 95% CI $[-19.89,3.47]$ (rough gloss vs. matte). That is, perceived focus decreased as surfaces exhibited more scattering in their reflectance function, but only in the background blur conditions. The effect of material in the background blur conditions suggests that the sharp ridge gradients generated by the more specular materials may have provided useful cues to the focus of the surface peaks. Most of these cues would have been destroyed by defocus in the foreground blur conditions, which may explain why the effect of material disappeared.

Figure 9

Download asset Open asset

Perceived gloss is depicted in Figure 10 and was analyzed with analogous contrasts to perceived focus. The layout of the plot is identical to Figure 9, but the vertical axis now represents the percentage of trials in which each image was selected as appearing glossier. The data indicate that foreground blur (right side) had a large negative effect on perceived gloss relative to the ‘no blur’ condition, ${\displaystyle t\left(9\right)=-21.57,p<0.001}$, 95% CI $[-42.09,-34.10]$, Cohen’s $d=7.19$. This effect was more severe for the materials with less scattering in their reflectance functions, which is likely because they were accurately perceived as being glossier in the no blur condition. The steep slopes for the smooth and rough gloss conditions indicate that optical defocus in the foreground rapidly destroyed the cues to gloss generated by the sharp surface ridge, but these large decrements in perceived gloss were not accompanied by large decrements in perceived focus. Together, these findings imply that observers partially misattributed the reduction in image focus in the foreground blur conditions to a change in surface material. Increasing background blur (left side) had a small significant negative effect on perceived gloss for the smooth gloss condition, $t\left(9\right)=-4.43,p=0.002$, 95% CI $[-12.95,-4.20]$, Cohen’s $d=1.48$, but no significant effect for the rough gloss ($t\left(9\right)=-0.098,p=0.924$, 95% CI [$-8.57,7.85]$]) or matte ($t\left(9\right)=2.52,p=0.033$, 95% CI $[0.917,16.940]$) conditions after the significance criterion was corrected with the Bonferroni method. This is likely due to the loss of gloss cues generated by the sharp specular reflections near the defocused base of the smooth glossy ridge.

Figure 10

Download asset Open asset

Cross-sectional depth profiles of perceived 3D shape constructed from the gauge figure settings are depicted in Figure 11. The profiles represent the average across observers after normalizing the mean height of each observer’s reconstructed profiles to the overall mean height. Differences in shape between the mean profiles were analyzed with an ANOVA in which each of the twenty probe positions was considered an independent sample; significant main effects of this ANOVA represent systematic changes in shape between pairs of focus conditions. The profiles for both the matte (top) and smooth gloss (bottom) conditions reveal that foreground blur had a significant effect on the perceived 3D shape of the ridge relative to the fully focused conditions, ${\displaystyle F\left(19,38\right)=34.48,p<0.001}$(matte) and ${\displaystyle F\left(19,38\right)=38.84,p<0.001}$ (smooth gloss). Background blur had no significant impact on perceived shape for either the matte surface, $F\left(\mathrm{19,38}\right)=1.42,p=1.78$, or the smooth gloss surface, $F\left(\mathrm{19,38}\right)=0.92,p=0.569$. The transformations in perceived 3D shape in the foreground blur condition involve a systematic reduction in ridge curvature, height, and position, which suggests that these perceptual distortions were not simply caused by a loss of information. As with perceived gloss, these findings indicate that observers partially misattributed the change in gradient structure induced by foreground blur to a reduction in surface curvature.

Figure 11

Download asset Open asset

Taken together, the results of the three tasks in Experiment 4 reveal that misperceptions of optical defocus are closely coupled with misperceptions of surface properties that also contribute to the smoothness of image gradients. The image features that appear to modulate these misperceptions in our stimuli are the sharp level cut contours that bound the surface ridge. When these contours were defocused by background blur, observers accurately reported a decrease in image focus, but when the ridge was defocused by foreground blur, the largest perceptual changes were instead in material (less gloss) and shape (less curvature). This suggests that the visual system was directly misattributing the smoothness of the image gradients in the foreground blur conditions to the wrong physical sources.

The demonstrations and experiments presented herein were designed to assess how the visual system disentangles the blurred, low-frequency image gradients generated by optical defocus from low-frequency gradients generated by focused shaded surfaces. The experiments were designed to assess the importance of photometric and geometric constraints in resolving this ambiguity. Our results revealed the existence of two types of misperceptions: illusory percepts of optical defocus when none is present (Experiments 1 and 3), and misperceptions of 3D shape and material when defocus is present but not detected (Experiment 4). We found that the perception of optical defocus arose in all shaded stimuli that lacked sharp bounding contours consistent with geometrically-correlated contours such as self-occlusions or level (planar) cuts of the surface. Taken together, our results indicate that this class of bounding contours plays a critical role in the modulating our experience of optical defocus and 3D shape in otherwise ambiguous images of shaded surfaces.

The main theoretical idea that shaped our experiments was that there are specific photogeometric constraints exhibited by the bounding contours of shaded surfaces that play a critical role in identifying low spatial frequency intensity gradients as surface shading, determining the contour’s border ownership, and establishing whether the surface is optically focused. We considered two primary constraints. The first was the covariation of contour orientation and shading intensity, which occurs generically for both smooth self-occluding contours and planar cuts (Figure 2A). The results of Experiment 1 confirm our informal observations that the rotated level cut mask (Figure 2B) not only failed to generate a clear perception of gradient focus, they actually interfered with the perception of 3D shape. We attribute this difference to the fact that the contours of level cut masks exhibit a strong orientation-intensity covariation, whereas the contours of the rotated masks do not.

We directly assessed the importance of this constraint in the perception of shading in Experiment 2, where thin shaded ‘ribbons’ tracing the paths of level cut contours were presented and the intensity-orientation covariation was directly manipulated. Our findings showed that this covariation strongly predicts the perception of 3D shape.

The second constraint we considered involved the direction of curvature of the shading adjacent to a contour (i.e., whether its perceived as convex or concave), and its role in the perceived attachment and focus of the shading gradients. The results of Experiment 1 support our informal observations that concavities appear less focused than convexities. This demonstrates that the intensity-orientation covariation cannot fully explain the perception of focus in images of shaded surfaces, as identical covariation can elicit percepts of both vivid focus and moderate defocus depending on whether the shaded surface is perceived as convex or concave (respectively). The difference in perceived focus may be caused by differences in perceived contour attachment in these two configurations. The shading gradients of the convex surface appear clearly attached to the contours of the level cut, but the same gradients do not appear clearly attached to the contour when the surface appears concave; the level cut contour can appear as the edge of a ‘cliff’, with the shading appearing at a more distant depth. This suggests that the shading gradients must appear clearly attached to the sharp contour to make full use of the information its sharpness provides about optical focus. This result is also consistent with arguments that the visual system has a bias to interpret abrupt discontinuities in luminance as self-occluding contours rather than sudden changes in surface orientation (Howard, 1983; Liu and Todd, 2004), which may explain why the contours appear more attached to the shading (and the shading more focused) when the surface appears convex.

Our findings have implications for the existing literature on both the perception of 3D shape from shading and the perception of optical blur. Prior studies of perceived shape have predominantly used globally convex shaded objects as stimuli (e.g. Fleming et al., 2004; Mingolla and Todd, 1986; Nefs et al., 2006). The self-occluding contours that are invariably exhibited by these stimuli may explain why their shading gradients are always perceived as fully focused and vividly three-dimensional. Our data also imply that attempts to ‘eliminate’ these self-occlusions will only impair perceived 3D shape to the extent that the orientation-intensity covariation exhibited in the image is actually reduced. Artificially cropping a self-occluding contour out of the image, for example, may simply create a new bounding contour that still exhibits enough covariation to induce percepts of focused 3D shading. This may explain why manipulations that rotate shaded terrains (which often exhibit no self-occlusions at all; Reichel and Todd, 1990; Todd and Reichel, 1989) have been found to negatively affect perceived shape more than cropping the self-occlusions of convex objects (Fleming et al., 2004; Egan and Todd, 2015). Our results suggest that manipulations involving planar cuts may be a more effective method of investigating the role of contours in surface perception in future work.

The literature on focus perception has predominantly investigated how the severity of perceived blur varies with the spatial frequency and contrast properties of images (Mather, 1997; O'Shea et al., 1997; Tadmor and Tolhurst, 1994). It has been established that relative changes in apparent blur magnitude can provide information about scene depth (Pentland, 1987) and that sharp bounding contours can resolve the depth ordering of ambiguous surfaces (Marshall et al., 1996), but to our knowledge, no prior studies have examined how the visual system distinguishes gradients produced by optical blur from gradients produced by environmental sources such as shading. Cases of source misattribution, such as those reported here, are likely difficult to produce with the 2D textures and simple contours employed in past work on focus perception. Our data reveal that identical low-frequency shading gradients can be perceived as vividly focused in some contexts and highly blurred in others, which implies that models of focus perception that rely entirely on local gradient features (or even the presence of sharp contours) are not sufficient. Our findings instead suggest that optical focus may be better characterized as a mid-level perceptual category that interacts with the visual system’s estimation of other mid-level properties such as contour attachment, 3D surface orientation and curvature, surface reflectance, and scene illumination.

The experiments and demonstrations reported herein have focused on the role of sharp contours that approximate smooth self-occluding rims in providing information about the 3D shape, depth of field, and optical focus of low frequency shading gradients. It seems unlikely, however, that sharp bounding contours are the sole means by which the visual system estimates the optical focus of shaded surfaces. Indeed, we carefully avoided other sources of image contours that could provide information about optical focus, such as the sharp contours generated by either shadows or specular reflections, and only evaluated low-curvature surfaces to avoid generated regions of high spatial frequency shading. Shaded surfaces with regions of high curvature could exhibit enough high spatial frequencies to eliminate percepts of blur, which could explain why some stimuli from prior studies of shading appear focused even in the absence of any contours (e.g. the ‘crater’ in Figure 7 of Todd et al., 2014). Specular reflections generated by low-curvature glossy surfaces also have similar spatial frequency properties to high-curvature shading (Mooney and Anderson, 2014). The efficacy of specular reflections in eliminating perceived blur can be experienced directly in the left panel of Figure 12. This surface is identical to that depicted in Figure 6 but was rendered in a natural light field with a specular reflectance component in addition to shading. This image appears as a fully focused, glossy, shaded surface. However, the perception of focus in this image requires that the specular reflections appear linked to the same surface geometry as the shading gradients; if the specular reflections are rotated to arbitrary positions, the shaded surface again appears blurred, and the specular reflections appear as overlaid pigment or a second, independent layer (right panel). Thus, specular reflections must also respect photogeometric constraints that link the specular and diffuse components of reflectance in order to provide information about optical focus, as has been shown previously in the perception of gloss (Anderson and Kim, 2009; Kim et al., 2011; Marlow et al., 2012). The underlying issue in both instances is source attribution: when the relevant image features (image highlights or sharp edges) are attributed to environmental sources (attached specular highlights or attached bounding contours), they simultaneously modulate the perception of environmental properties (gloss or 3D shape) and optical focus. Note that, in general, realistic specular reflections are not optically attached to the surface; when the depth of field is reduced, different regions of specular reflections may only remain sharp when viewed with different focal lengths, and may not be in focus at the same focal length as the attached surface shading. The interactions between specular reflectance, depth of field, and perceived focus are worth consideration in future work.

Figure 12

Download asset Open asset

The demonstrations and experiments presented here have focused on a previously unappreciated computational problem: distinguishing low-frequency structure caused by optical blur from the shading gradients of smooth surfaces. We showed that the visual system exploits specific forms of geometric and photometric covariation to generate percepts of optical focus and vivid surface shading, which arise generically along both smooth self-occlusions and planar cuts. Our results indicate that the presence or absence of intensity-orientation correlated contours is a powerful cue to focus: fully focused shaded surfaces can appear blurred in their absence, and actual defocus can be mistaken for transformations in material and shape when correlated contours are nearby. The findings reported herein provide evidence that our perceptions of material, 3D shape, and optical defocus are inherently coupled, which suggests that optical defocus perception does not occur in a completely independent visual pathway to the perception of surface and scene properties. Future work is required to understand the neural processes that exploit these sources of covariation, and the other sources of information that the visual system utilizes to distinguish environmental sources of image structure from optical artifacts induced by the imaging properties of single-chambered eyes.

All data generated or analysed during this study are included in the manuscript and supporting files. The code for ribbon generation and correlation computation is available on GitHub (https://github.com/swjmooney/contour-covariation; copy archived at https://github.com/elifesciences-publications/contour-covariation).

1. 1. Anderson BL
   2. Kim J

   (2009) Image statistics do not explain the perception of gloss and lightness

   Journal of Vision 9:10.

   https://doi.org/10.1167/9.11.10

   • PubMed
   • Google Scholar
2. 1. Barrow H
   2. Tenenbaum J

   (1978)

   Recovering intrinsic scene characteristics

   Comput Vis Syst 2:3–26.

   • Google Scholar
3. 1. Belhumeur PN
   2. Kriegman DJ
   3. Yuille AL

   (1999)

   The bas-relief ambiguity

   International Journal of Computer Vision 35:33–44.

   • Google Scholar
4. 1. Ciuffreda KJ
   2. Selenow A
   3. Wang B
   4. Vasudevan B
   5. Zikos G
   6. Ali SR

   (2006) "Bothersome blur": a functional unit of blur perception

   Vision Research 46:895–901.

   https://doi.org/10.1016/j.visres.2005.10.004

   • PubMed
   • Google Scholar
5. 1. Crete F
   2. Dolmiere T
   3. Ladret P
   4. Nicolas M

   (2007) The blur effect: perception and estimation with a new no-reference perceptual blur metric

   Proceedings of SPIE - the International Society for Optical Engineering.

   https://doi.org/10.1117/12.702790

   • Google Scholar
6. 1. Egan EJ
   2. Todd JT

   (2015) The effects of smooth occlusions and directions of illumination on the visual perception of 3-D shape from shading

   Journal of Vision 15:24.

   https://doi.org/10.1167/15.2.24

   • PubMed
   • Google Scholar
7. Conference

   1. Ferzli R
   2. Karam LJ

   (2006) Human visual system based no-reference objective image sharpness metric

   2006 International Conference on Image Processing. pp. 2949–2952.

   https://doi.org/10.1109/ICIP.2006.312925

   • Google Scholar
8. 1. Fleming RW
   2. Dror RO
   3. Adelson EH

   (2003) Real-world illumination and the perception of surface reflectance properties

   Journal of Vision 3:3.

   https://doi.org/10.1167/3.5.3

   • Google Scholar
9. 1. Fleming RW
   2. Torralba A
   3. Adelson EH

   (2004) Specular reflections and the perception of shape

   Journal of Vision 4:10.

   https://doi.org/10.1167/4.9.10

   • Google Scholar
10. Conference

    1. Halstead M
    2. Kass M
    3. DeRose T

    (1993) Efficient, fair interpolation using Catmull-Clark surfaces

    Proceedings of the 20th Annual Conference on Computer Graphics and Interactive Techniques. pp. 35–44.

    https://doi.org/10.1145/166117.166121

    • Google Scholar
11. 1. Held RT
    2. Cooper EA
    3. O'Brien JF
    4. Banks MS

    (2010) Using blur to affect perceived distance and size

    ACM Transactions on Graphics 29:1–16.

    https://doi.org/10.1145/1731047.1731057

    • PubMed
    • Google Scholar
12. 1. Hill H
    2. Bruce V

    (1994) A comparison between the hollow-face and 'hollow-potato' illusions

    Perception 23:1335–1337.

    https://doi.org/10.1068/p231335

    • PubMed
    • Google Scholar
13. Book

    1. Horn BK
    2. Brooks MJ

    (1989)

    Shape From Shading

    MIT Press.

    • Google Scholar
14. 1. Howard IP

    (1983) Occluding edges in apparent reversal of convexity and concavity

    Perception 12:85–86.

    https://doi.org/10.1068/p120085

    • PubMed
    • Google Scholar
15. 1. Kim J
    2. Marlow P
    3. Anderson BL

    (2011) The perception of gloss depends on highlight congruence with surface shading

    Journal of Vision 11:4.

    https://doi.org/10.1167/11.9.4

    • PubMed
    • Google Scholar
16. 1. Koenderink JJ

    (1984) What does the occluding contour tell us about solid shape?

    Perception 13:321–330.

    https://doi.org/10.1068/p130321

    • PubMed
    • Google Scholar
17. 1. Koenderink JJ
    2. van Doorn AJ
    3. Kappers AM

    (1992) Surface perception in pictures

    Perception & Psychophysics 52:487–496.

    https://doi.org/10.3758/BF03206710

    • PubMed
    • Google Scholar
18. 1. Koenderink JJ
    2. van Doorn AJ
    3. Kappers AM
    4. Todd JT

    (2001) Ambiguity and the 'mental eye' in pictorial relief

    Perception 30:431–448.

    https://doi.org/10.1068/p3030

    • PubMed
    • Google Scholar
19. 1. Koenderink JJ
    2. van Doorn AJ
    3. Pont SC

    (2004) Light direction from shad(ow)ed random gaussian surfaces

    Perception 33:1405–1420.

    https://doi.org/10.1068/p5287

    • PubMed
    • Google Scholar
20. 1. Koenderink JJ
    2. Van Doorn AJ
    3. Pont SC

    (2007) Perception of illuminance flow in the case of anisotropic rough surfaces

    Perception & Psychophysics 69:895–903.

    https://doi.org/10.3758/BF03193926

    • PubMed
    • Google Scholar
21. 1. Langer MS
    2. Bülthoff HH

    (2001) A prior for global convexity in local shape-from-shading

    Perception 30:403–410.

    https://doi.org/10.1068/p3178

    • PubMed
    • Google Scholar
22. 1. Langer MS
    2. Zucker SW

    (1994) Shape-from-shading on a cloudy day

    Journal of the Optical Society of America A 11:467–478.

    https://doi.org/10.1364/JOSAA.11.000467

    • Google Scholar
23. 1. Liu B
    2. Todd JT

    (2004) Perceptual biases in the interpretation of 3D shape from shading

    Vision Research 44:2135–2145.

    https://doi.org/10.1016/j.visres.2004.03.024

    • PubMed
    • Google Scholar
24. 1. Mamassian P

    (1993)

    Geometric Methods in Computer Vision II

    124–133, Isophotes on a smooth surface related to scene geometry, Geometric Methods in Computer Vision II, International Society for Optics and Photonics Vol. 2031.

    • Google Scholar
25. 1. Marlow PJ
    2. Kim J
    3. Anderson BL

    (2012) The perception and misperception of specular surface reflectance

    Current Biology 22:1909–1913.

    https://doi.org/10.1016/j.cub.2012.08.009

    • PubMed
    • Google Scholar
26. 1. Marlow PJ
    2. Mooney SWJ
    3. Anderson BL

    (2019) Photogeometric cues to perceived surface shading

    Current Biology 29:306–311.

    https://doi.org/10.1016/j.cub.2018.11.041

    • PubMed
    • Google Scholar
27. 1. Marshall JA
    2. Burbeck CA
    3. Ariely D
    4. Rolland JP
    5. Martin KE

    (1996) Occlusion edge blur: a cue to relative visual depth

    Journal of the Optical Society of America A 13:681–688.

    https://doi.org/10.1364/JOSAA.13.000681

    • PubMed
    • Google Scholar
28. 1. Mather G

    (1996) Image blur as a pictorial depth cue

    Proceedings. Biological Sciences 263:169–172.

    https://doi.org/10.1098/rspb.1996.0027

    • PubMed
    • Google Scholar
29. 1. Mather G

    (1997) The use of image blur as a depth cue

    Perception 26:1147–1158.

    https://doi.org/10.1068/p261147

    • PubMed
    • Google Scholar
30. 1. Mather G
    2. Smith DR

    (2000) Depth cue integration: stereopsis and image blur

    Vision Research 40:3501–3506.

    https://doi.org/10.1016/S0042-6989(00)00178-4

    • PubMed
    • Google Scholar
31. 1. Mather G
    2. Smith DR

    (2002) Blur discrimination and its relation to blur-mediated depth perception

    Perception 31:1211–1219.

    https://doi.org/10.1068/p3254

    • PubMed
    • Google Scholar
32. 1. Mingolla E
    2. Todd JT

    (1986) Perception of solid shape from shading

    Biological Cybernetics 53:137–151.

    https://doi.org/10.1007/BF00342882

    • PubMed
    • Google Scholar
33. Software

    1. Mooney SWJ

    (2019) contour-covariation, version 4e49b20

    GitHub.

    https://github.com/swjmooney/contour-covariation
34. 1. Mooney SW
    2. Anderson BL

    (2014) Specular image structure modulates the perception of three-dimensional shape

    Current Biology 24:2737–2742.

    https://doi.org/10.1016/j.cub.2014.09.074

    • PubMed
    • Google Scholar
35. 1. Nefs HT
    2. Koenderink JJ
    3. Kappers AM

    (2006) Shape-from-shading for matte and glossy objects

    Acta Psychologica 121:297–316.

    https://doi.org/10.1016/j.actpsy.2005.08.001

    • PubMed
    • Google Scholar
36. 1. O'Shea RP
    2. Govan DG
    3. Sekuler R

    (1997) Blur and contrast as pictorial depth cues

    Perception 26:599–612.

    https://doi.org/10.1068/p260599

    • PubMed
    • Google Scholar
37. Conference

    1. Oren M
    2. Nayar SK

    (1993) Diffuse reflectance from rough surfaces

    Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. pp. 763–764.

    https://doi.org/10.1109/CVPR.1993.341163

    • Google Scholar
38. Conference

    1. Pellacini F
    2. Ferwerda JA
    3. Greenberg DP

    (2000) Toward a psychophysically-based light reflection model for image synthesis

    Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques. pp. 55–64.

    https://doi.org/10.1145/344779.344812

    • Google Scholar
39. 1. Pentland AP

    (1984) Local shading analysis

    IEEE Transactions on Pattern Analysis and Machine Intelligence 6:170–187.

    https://doi.org/10.1109/TPAMI.1984.4767501

    • PubMed
    • Google Scholar
40. 1. Pentland AP

    (1987) A new sense for depth of field

    IEEE Transactions on Pattern Analysis and Machine Intelligence 9:523–531.

    https://doi.org/10.1109/TPAMI.1987.4767940

    • PubMed
    • Google Scholar
41. 1. Ramachandran VS

    (1988) Perception of shape from shading

    Nature 331:163–166.

    https://doi.org/10.1038/331163a0

    • PubMed
    • Google Scholar
42. 1. Reichel FD
    2. Todd JT

    (1990) Perceived depth inversion of smoothly curved surfaces due to image orientation

    Journal of Experimental Psychology: Human Perception and Performance 16:653–664.

    https://doi.org/10.1037/0096-1523.16.3.653

    • Google Scholar
43. 1. Tadmor Y
    2. Tolhurst DJ

    (1994) Discrimination of changes in the second-order statistics of natural and synthetic images

    Vision Research 34:541–554.

    https://doi.org/10.1016/0042-6989(94)90167-8

    • PubMed
    • Google Scholar
44. 1. Todd JT
    2. Koenderink JJ
    3. van Doorn AJ
    4. Kappers AM

    (1996) Effects of changing viewing conditions on the perceived structure of smoothly curved surfaces

    Journal of Experimental Psychology: Human Perception and Performance 22:695–706.

    https://doi.org/10.1037/0096-1523.22.3.695

    • PubMed
    • Google Scholar
45. 1. Todd JT
    2. Egan EJ
    3. Phillips F

    (2014) Is the perception of 3D shape from shading based on assumed reflectance and illumination?

    I-Perception 5:497–514.

    https://doi.org/10.1068/i0645

    • PubMed
    • Google Scholar
46. 1. Todd JT
    2. Reichel FD

    (1989) Ordinal structure in the visual perception and cognition of smoothly curved surfaces

    Psychological Review 96:643–657.

    https://doi.org/10.1037/0033-295X.96.4.643

    • PubMed
    • Google Scholar
47. 1. Watt SJ
    2. Akeley K
    3. Ernst MO
    4. Banks MS

    (2005) Focus cues affect perceived depth

    Journal of Vision 5:834–862.

    https://doi.org/10.1167/5.10.7

    • Google Scholar
48. 1. Webster MA
    2. Georgeson MA
    3. Webster SM

    (2002) Neural adjustments to image blur

    Nature Neuroscience 5:839–840.

    https://doi.org/10.1038/nn906

    • PubMed
    • Google Scholar

Author details

1. Scott WJ Mooney

   School of Psychology, The University of Sydney, Sydney, Australia

   Contribution

   Conceptualization, Software, Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review and editing

   For correspondence

   scm2011@med.cornell.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-0094-7638

2. Phillip J Marlow

   School of Psychology, The University of Sydney, Sydney, Australia

   Contribution

   Conceptualization, Writing—review and editing

   Competing interests

   No competing interests declared

3. Barton L Anderson

   School of Psychology, The University of Sydney, Sydney, Australia

   Contribution

   Supervision, Investigation, Writing—review and editing

   Competing interests

   No competing interests declared

• 1,573

  views

• 133

  downloads

• 9

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.