Viewpoint-dependent recognition of scenes

Three experiments investigated scene recognition across viewpoint changes, involving same/different judgements on scenes consisting of three objects on a desktop. On same trials, the comparison scene appeared either from the same viewpoint as the standard scene or from a different viewpoint with the desktop rotated about one or more axes. Different trials were created either by interchanging the locations of two or three of the objects (location change condition), or by rotating either one or all three of the objects around their vertical axes (orientation change condition). Response times and errors increased as a function of the angular distance between the standard and comparison views, but this effect was bigger for rotations around the vertical axis than for those about the line of sight or horizontal axis. Furthermore, the time to detect location changes was less than that to detect orientation changes, and this difference increased with increasing angular disparity between the standard and comparison scenes. Rotation times estimated in a double-axis rotation were no longer than other rotations in depth, indicating that alignment was not necessarily simpler around a "natural" axis of rotation. These results are consistent with the hypothesis that scenes, like many objects, may be represented in a viewpoint dependent manner and recognized by aligning standard and comparison views, but that the alignment of scenes is not a holistic process.     

Use of preferential inspection to define the viewing sphere and characteristic views of an arbitrary machined tool part.

Measurements were made of the way human subjects visually inspected an idealized machined tool part (a 'widget') while learning the three-dimensional shape of the object. Subjects were free to rotate the object about any axis. Inspection was not evenly distributed across all views. Subjects focused on views where the faces of the object were orthogonal to the line of sight and the edges of the object were aligned parallel or at right angles to the gravitational axis. These 'face' or 'plan' views were also the easiest for subjects to bring to mind in a mental imagery task. By contrast, when subjects were instructed to imagine the views displaying the most structural information they visualized views lying midway between face views.     

Mental imagery and the third dimension

What sort of medium underlies imagery for three-dimensional scenes? In the present investigation, the time subjects took to scan between objects in a mental image was used to infer the sorts of geometric information that images preserve. Subjects studied an open box in which five objects were suspended, and learned to imagine this display with their eyes closed. In the first experiment, subjects scanned by tracking an imaginary point moving in a straight line between the imagined objects. Scanning times increased linearly with increasing distance between objects in three dimensions. Therefore metric 3-D information must be preserved in images, and images cannot simply be 2-D "snapshots." In a second experiment, subjects scanned across the image by "sighting" objects through an imaginary rifle sight. Here scanning times were found to increase linearly with the two-dimensional separations between objects as they appeared from the original viewing angle. Therefore metric 2-D distance information in the original perspective view must be preserved in images, and images cannot simply be 3-D "scale-models" that are assessed from any and all directions at once. In a third experiment, subjects mentally rotated the display 90 degrees and scanned between objects as they appeared in this new perspective view by tracking an imaginary rifle signt, as before. Scanning times increased linearly with the two-dimensional separations between objects as they would appear from the new relative viewing perspective. Therefore images can display metric 2-D distance information in a perspective view never actually experiences, so mental images cannot simply be "snapshot plus scale model" pairs. These results can be explained by a model in which the three-dimensional structure of objects is encoded in long-term memory in 3-D object-centered coordinate systems. When these objects are imagined, this information is then mapped onto a single 2-D "surface display" in which the perspective properties specific to a given viewing angle can be depicted. In a set of perceptual control experiments, subjects scanned a visible display by (a) simply moving their eyes from one object to another, (b) sweeping an imaginary rifle sight over the display, or (c) tracking an imaginary point moving from one object to another. Eye-movement times varied linearly with 2-D interobject distance, as did time to scan with an imaginary rifle sight; time to tract a point varied independently with the 3-D and 2-D interobject distances. These results are compared with the analogous image scanning results to argue that imagery and perception share some representational structures but that mental image scanning is a process distinct from eye movements or eye-movement commands.     

Recognizing a plane-misoriented view of a familiar object is not influenced by the ease of specifying the main axis of elongation of that object

Participants saw three versions of pictures of familiar objects: the original unaltered (axis-normal) pictures, axis-extended pictures in which the main axes of the axis-normal pictures were elongated, and axis-switched pictures in which objects that were originally horizontally elongated were depicted as vertically elongated and vice versa. Relative to axis-normal pictures, axis extension aided decisions about whether the picture of the object was wide or tall, and axis switching hindered these decisions for both upright and plane-misoriented views. Nevertheless, although these axis manipulations clearly influenced decisions about the location of the object's main axis of elongation, axis-switched pictures were no harder to name than axis-extended pictures. Changing the depicted main axis of elongation by axis switching and axis extension did not influence object recognition in itself, whether for upright or for plane-misoriented views. This suggests that specifying the main axis of elongation of an object does not play an important role in the orientation-sensitive processes involved in identifying plane-misoriented views of that object.     

Visual representation of malleable and rigid objects that deform as they rotate

Most studies and theories of object recognition have addressed the perception of rigid objects. Yet, physical objects may also move in a nonrigid manner. A series of priming studies examined the conditions under which observers can recognize novel views of objects moving nonrigidly. Observers were primed with 2 views of a rotating object that were linked by apparent motion or presented statically. The apparent malleability of the rotating prime object varied such that the object appeared to be either malleable or rigid. Novel deformed views of malleable objects were primed when falling within the object's motion path. Priming patterns were significantly more restricted for deformed views of rigid objects. These results suggest that moving malleable objects may be represented as continuous events, whereas rigid objects may not. That is, object representations may be "dynamically remapped" during the analysis of the object's motion.     

Viewpoint and orientation influence picture recognition in the blind

In the first three experiments, subjects felt solid geometrical forms and matched raised-line pictures to the objects. Performance was best in experiment 1 for top views, with shorter response latencies than for side views, front views, or 3-D views with foreshortening. In a second experiment with blind participants, matching accuracy was not significantly affected by prior visual experience, but speed advantages were found for top views, with 3-D views also yielding better matching accuracy than side views. There were no performance advantages for pictures of objects with a constant cross section in the vertical axis. The early-blind participants had lower performance for side and frontal views. The objects were rotated to oblique orientations in experiment 3. Early-blind subjects performed worse than the other subjects given object rotation. Visual experience with pictures of objects at many angles could facilitate identification at oblique orientations. In experiment 5 with blindfolded sighted subjects, tangible pictures were used as targets and as choices. The results yielded superior overall performance for 3-D views (mean, M = 74% correct) and much lower matching accuracy for top views as targets (M = 58% correct). Performance was highest when the target and matching viewpoint were identical, but 3-D views (M = 96% correct) were still far better than top views. The accuracy advantage of the top views also disappeared when more complex objects were tested in experiment 6. Alternative theoretical implications of the results are discussed.     

Visual search for object orientation can be modulated by canonical orientation

The authors studied the influence of canonical orientation on visual search for object orientation. Displays consisted of pictures of animals whose axis of elongation was either vertical or tilted in their canonical orientation. Target orientation could be either congruent or incongruent with the object's canonical orientation. In Experiment 1, vertical canonical targets were detected faster when they were tilted (incongruent) than when they were vertical (congruent). This search asymmetry was reversed for tilted canonical targets. The effect of canonical orientation was partially preserved when objects were high-pass filtered, but it was eliminated when they were low-pass filtered, rendering them as unfamiliar shapes (Experiment 2). The effect of canonical orientation was also eliminated by inverting the objects (Experiment 3) and in a patient with visual agnosia (Experiment 4). These results indicate that orientation search with familiar objects can be modulated by canonical orientation, and they indicate a top-down influence on orientation processing.     

Seeing big things: overestimation of heights is greater for real objects than for objects in pictures

In six experiments we demonstrate that the vertical-horizontal illusion that is evoked when viewing photographs and line drawings is relatively small, whereas the magnitude of this illusion when large objects are viewed is at least twice as great. Furthermore, we show that the illusion is due more to vertical overestimation than horizontal underestimation. The lack of a difference in vertical overestimation between pictures and line drawings suggests that vertical overestimation in pictures depends solely on the perceived physical size of the projection on the picture surface, rather than on what is apparent about an object's represented size. The vertical-horizontal illusion is influenced by perceived physical size. It is greater when viewing large objects than small pictures of these same objects, even when visual angles are equated.     

Young Children's Self-Generated Object Views and Object Recognition

Two important and related developments in children between 18 and 24 months of age are the rapid expansion of object name vocabularies and the emergence of an ability to recognize objects from sparse representations of their geometric shapes. In the same period, children also begin to show a preference for planar views (i.e., views of objects held perpendicular to the line of sight) of objects they manually explore. Are children's emerging view preferences somehow related to contemporary changes in object name vocabulary and object perception? Children aged 18 to 24 months old explored richly detailed toy objects while wearing a head camera that recorded their object views. Both children's vocabulary size and their success in recognizing sparse three-dimensional representations of the geometric shapes of objects were significantly related to their spontaneous choice of planar views of those objects during exploration. The results suggest important interdependencies among developmental changes in perception, action, word learning, and categorization in very young children.     

Some views are better than others: evidence for a visual bias in object views self‐generated by toddlers

How objects are held determines how they are seen, and may thereby play an important developmental role in building visual object representations. Previous research suggests that toddlers, like adults, show themselves a disproportionate number of planar object views – that is, views in which the objects' axes of elongation are perpendicular or parallel to the line of sight. Here, three experiments address three explanations of this bias: (1) that the locations of interesting features of objects determine how they are held and thus how they are viewed; (2) that ease of holding determines object views; and (3) that there is a visual bias for planar views that exists independently of holding and of interesting surface properties. Children 18 to 24 months of age manually and visually explored novel objects (1) with interesting features centered in planar or ¾ views; (2) positioned inside Plexiglas boxes so that holding biased either planar or non‐planar views; and (3) positioned inside Plexiglas spheres, so that no object properties directly influenced holding. Results indicate a visual bias for planar views that is influenced by interesting surface properties and ease of holding, but that continues to exist even when these factors push for alternative views.     

Effects of texture and shape on perceived time to passage: knowing "what" influences judging "when"

In the present study, we examined whether it is easier to judge when an object will pass one's head if the object's surface is textured. There are three reasons to suspect that this might be so: First, the additional (local) optic flow may help one judge the rate of expansion and the angular velocity more reliably. Second, local deformations related to the change in angle between the object and the observer could help track the object's position along its path. Third, more reliable judgments of the object's shape could help separate global expansioncaused by changes in distance from expansion due to changes in the angle between the object and the observer. We can distinguish among these three reasons by comparing performance for textured and uniform spheres and disks. Moving objects were displayed for 0.5-0.7 sec. Subjects had to decide whether the object would pass them before or after a beep that was presented 1 sec after the object started moving. Subjects were not more precise with textured objects. When the disk rotated in order to compensate for the orientation-related contraction that its image would otherwise undergo during its motion, it appeared to arrive later, despite the fact that this strategy increases the global rate of expansion. We argue that this is because the expected deformation of the object's image during its motion is considered when time to passage is judged. Therefore, the most important role for texture in everyday judgments of time to passage is probably that it helps one judge the object's shape and thereby estimate how its image will deform as it moves.     

Visual angle as a determinant of perceived interobject distance

Two new experiments and a reanalysis of Toye’s (1986)data are used to examine the relationship between true distance and perceived distance in natural scenes. In the first experiment, 8 subjects estimated 78 interobject distances, formed by all pairs of 13 objects, while viewing the objects from a fixed position. The results showed that estimated distance is a linear function of the visual angle between objects as well as of the true distance. This relationship results in distances perpendicular to the line of sight being overestimated in relation to true distances and to distances parallel to the line of sight. These findings were confirmed by reanalysis of a comparable data set from Toye. Since changes in the visual angle can come about through changes in alignment with the line of sight, viewing distance, or interobject distance, Experiment 2 was designed to determine whether the visual angle effect was due to one of these, or whether it was an independent effect. In Experiment 2, 8 subjects estimated six interobject distances from 12 viewing positions. The results showed that visual angle predicted estimated distance independently of how the change in visual angle came about, suggesting that the greater the visual angle between objects, the more their separation is overestimated.     

The effects of size,clutter, and complexity on vanishing-point distances in visual imagery

Summary The portrayal of vanishing-point distances in visual imagery was examined in six experiments. In all experiments, subjects formed visual images of squares, and the squares were to be oriented orthogonally to subjects' line of sight. The squares differed in their level of surface complexity, and were either undivided, divided into 4 equally sized smaller squares, or divided into 16 equally sized smaller squares. Squares also differed in stated referent size, and ranged from 3 in. to 128 ft along each side. After subjects had formed an image of a specified square, they transformed their image so that the square was portrayed to move away from them. Eventually, the imaged square was portrayed to be so far away that if it were any further away, it could not be identified. Subjects estimated the distance to the square that was portrayed in their image at that time, the vanishing-point distance, and the relationship between stated referent size and imaged vanishing-point distance was best described by a power function with an exponent less than 1. In general, there were trends for exponents (slopes on log axes) to increase slightly and for multiplicative constants (y intercepts on log axes) to decrease as surface complexity increased. No differences in exponents or in multiplicative constants were found when the vanishing-point was approached from either sub-threshold or suprathreshold directions. When clutter in the form of additional imaged objects located to either side of the primary imaged object was added to the image, the exponent of the vanishing-point function increased slightly and the multiplicative constant decreased. The success of a power function (and the failure of the size-distance invariance hypothesis) in describing the vanishing-point distance function calls into question the notions (a) that a constant grain size exists in the, imaginal visual field at a given location and (b) that grain size specifies a lower limit in the storage of information in visual images.     

A corpus of 714 full-color images of depth-rotated objects

A set of full-color images of objects is described for use in experiments investigating the effects of in-depth rotation on the identification of three-dimensional objects. The corpus contains up to 11 perspective views of 70 nameable objects. We also provide ratings of the “goodness” of each view, based on Thurstonian scaling of subjects’ preferences in a paired-comparison experiment. An exploratory cluster analysis on the scaling solutions indicates that the amount of information available in a given view generally is the major determinant of the goodness of the view. For instance, objects with an elongated front-back axis tend to cluster together, and the front and back views of these objects, which do not reveal the object’s major surfaces and features, are evaluated as the worst views.     

Aesthetic issues in spatial composition: effects of vertical position and perspective on framing single objects

Aesthetic preference for the vertical composition of single-object pictures was studied through a series of two-alternative forced-choice experiments. The results reveal the influence of several factors, including spatial asymmetries in the functional properties of the object and the typical position of the object relative to the observer. With asymmetric side views of objects, people generally prefer objects typically located below the observer's viewpoint (e.g., a bowl or swimming stingray) to be below the center of the frame and objects typically located above the observer's viewpoint (e.g., a light fixture or flying eagle) to be above the center of the frame. In addition, people generally prefer symmetric views of those same objects from directly above or directly below to be closer to the center of the frame. We suggest that these results can be unified by the hypothesis that people prefer the object's "affordance space" to be centered within the frame.     

Interactions between symmetry and elongation in determining reference frames for object perception.

Many theories of object recognition posit that objects are encoded with respect to a perceptual frame of reference. Such theories assume that factors such as symmetry and elongation are critical for the assignment of an object's primary axis, and consequently for the extraction of an object's reference frame. The present experiments directly examined the relative roles played by symmetry and elongation in the determination of an object's primary axis, and the extent to which symmetry and elongation interact with one another. We found that observers use both symmetry and elongation in extracting an object's primary axis, that the extent to which each cue dominates depends on its relative salience, and that symmetry and elongation are processed interactively, rather than in encapsulated modules.     

Viewer-external frames of reference in the mental transformation of 3-D objects

Most models of object recognition and mental rotation are based on the matching of an object's 2-D view with representations of the object stored in memory. They propose that a time-consuming normalization process compensates for any difference in viewpoint between the 2-D percept and the stored representation. Our experiment shows that such normalization is less time consuming when it has to compensate for disorientations around the vertical than around the horizontal axis of rotation. By decoupling the different possible reference frames, we demonstrate that this anisotropy of the normalization process is defined not with respect to the retinal frame of reference, but, rather, according to the gravitational or the visuocontextual frame of reference. Our results suggest that the visual system may call upon both the gravitational vertical and the visuocontext to serve as the frame of reference with respect to which 3-D objects are gauged in internal object transformations.     

The perceptual analysis of structure from motion for rotating objects undergoing affine stretching transformations

In two experiments, we evaluated the ability of human observers to make use of second-order temporal relations across three or more views of an apparent motion sequence for the perceptual analysis of three-dimensional form. Ratings of perceived rigidity were obtained in Experiment 1 for objects rotating in depth that were simultaneously subjected to sinusoidal affine stretching transformations along the line of sight or in a direction parallel to the image plane. Such transformations are theoretically interesting because they cannot be detected by analyses that are restricted to first-order temporal relations (i.e., two views), but they can be detected by more conventional analyses of structure from motion in which second-order temporal relations over three or more views are used. The current results show that human observers can perceive stretching transformations of a rotating 3-D object in a direction parallel to the image plane but that they fail to perceive stretching transformations along the line of sight. This result suggests that human observers can make use of some limited second-order temporal information. This finding was confirmed in Experiment 2, in which we investigated the effects of several specific optical consequences of sinusoidal stretching transformations applied in different directions. The results indicate that observers may be sensitive to the sign of acceleration, but that. they cannot make use of the precise magnitude of second-order relations necessary to recover euclidean metric structure.     

Seeing two sides at once: effects of viewpoint and object structure on recognizing three-dimensional objects

Five experiments demonstrated that adults can identify certain novel views of 3-dimensional model objects on the basis of knowledge of a single perspective. Geometrically irregular contour (wire) and surface (clay) objects and geometrically regular surface (pipe) objects were accurately recognized when rotated 180 degrees about the vertical (y) axis. However, recognition accuracy was poor for all types of objects when rotated around the y-axis by 90 degrees. Likewise, more subtle rotations in depth (i.e., 30 degrees and 60 degrees) induced decreases in recognition of both contour and surface objects. These results suggest that accurate recognition of objects rotated in depth by 180 degrees may be achieved through use of information in objects' 2-dimensional bounding contours, the shapes of which remain invariant over flips in depth. Consistent with this interpretation, a final study showed that even slight rotations away from 180 degrees cause precipitous drops in recognition accuracy.     

Three gradients and the perception of flat and curved surfaces

Researchers of visual perception have long been interested in the perceived slant of a surface and in the gradients that purportedly specify it. Slant is the angle between the line of sight and the tangent to the planar surface at any point, also called the surface normal. Gradients are the sources of information that grade, or change, with visual angle as one looks from one's feet upward to the horizon. The present article explores three gradients--perspective, compression, and density--and the phenomenal impression of flat and curved surfaces. The perspective gradient is measured at right angles to the axis of tilt at any point in the optic array; that is, when looking down a hallway at the tiles of a floor receding in the distance, perspective is measured by the x-axis width of each tile projected on the image plane orthogonal to the line of sight. The compression gradient is the ratio of y/x axis measures on the projected plane. The density gradient is measured by the number of tiles per unit solid visual angle. For flat surfaces and many others, perspective and compression gradients decrease with distance, and the density gradient increases. We discuss the manner in which these gradients change for various types of surfaces. Each gradient is founded on a different assumption about textures on the surfaces around us. In Experiment 1, viewers assessed the three-dimensional character of projections of flat and curved surfaces receding in the distance. They made pairwise judgments of preference and of dissimilarity among eight stimuli in each of four sets. The presence of each gradient was manipulated orthogonally such that each stimulus had zero, one, two, or three gradients appropriate for either a flat surface or a curved surface. Judgments were made were made for surfaces with both regularly shaped and irregularly shaped textures scattered on them. All viewer assessment were then scaled in one dimension. Multiple correlation and regression on the scale values revealed that greater than 98% of the variance in scale values was accounted for by the gradients. For the flat surfaces a mean of 65% of the variance was accounted for by the perspective gradient, 28% by the density gradient, and 6% by the compression gradient. For curved surfaces, on the other hand, a mean of 96% of the variance was accounted for by the compression gradient, and less than 2% by either the perspective gradient or the density gradient.(ABSTRACT TRUNCATED AT 400 WORDS)