Search

One of the major difficulties in graph classification is the lack of mathematical structure in the space of graphs. The use of kernel machines allows us to overcome this fundamental limitation in an elegant manner by addressing the pattern recognition problem in an implicitly existing feature vector space instead of the original space of graphs. In this paper we propose three novel error-tolerant graph kernels -- a diffusion kernel, a convolution kernel, and a random walk kernel. The kernels are closely related to one of the most flexible graph matching methods, graph edit distance. Consequently, our kernels are applicable to virtually any kind of graph. They also show a high degree of robustness against various types of distortion. In an experimental evaluation involving the classification of line drawings, images, diatoms, fingerprints, and molecules, we demonstrate the superior performance of the proposed kernels in conjunction with support vector machines over a standard nearest-neighbor reference method and several other graph kernels including a standard random walk kernel.

This paper describes three examples of illusion installation in a science museum environment from the author's collaboration with the artist and architect. The installations amplify the illusory effects, such as vection (visually-induced sensation of self motion) and motion-induced blindness, to emphasize that perception is not just to obtain structure and features of objects, but rather to grasp the dynamic relationship between the self and the world. Scaling up the size and utilizing the live human body turned out to be keys for installations with higher emotional impact.

This paper is a call for greater use of individual differences in the basic science of visual perception. Individual differences yield insights into visual perception’s functional organization, underlying biological/environmental mechanisms, and utility. I first explain the general approach advocated and where it comes from. Second, I describe five principles central to learning about the nature of visual perception through individual differences. Third, I elaborate on the use of individual differences to gain insights into the three areas mentioned above (function, biology/environment, utility), in each case describing the approach advocated, presenting model examples from the literature, and laying out illustrative research proposals for the case of stereopsis.

Rules of composition in paintings form a rich probe into the principles of perceptual processing that have been discussed for centuries. These principles can be studied by controlled scientific experiments, but an alternative approach is to use the art works themselves as a database for direct analysis. This paper focuses on the analysis of composition in relation to the canvas frame. An underlying principle is the compositional pyramid rising from the bottom of the frame to a center of consciousness high on the midline, which also finds its expression in the configuration of portrait paintings. The analyses presented reveal a dominant positioning principle for one eye in a portrait to lie on the vertical axis with an unbiased accuracy of the order of +/-5%. Analysis of the vertical location shows that the dominant height is at or above the Golden Section level on the vertical axis. In general, the layout of the portrait follows the principle of the compositional pyramid, with a center of consciousness at its apex, but there are many other compositional principles at work in the corpus of portraits in general. Analysis of the portraits of particular artists reveals that special features of their work must be considered in order to identify those that do and do not conform to the eye-centering principle.

When presented against a highly lit black background, dimly illuminated white paper strips appear white even when they are equiluminant with the background. Such an example of simultaneous lightness constancy cannot be accounted for by receptor gain control because of the equiluminance. Moreover, this demonstration shows that lightness cannot be reduced to 'relative brightness' as is widely believed.

Beck (1972) pointed out that grouping tasks, where one set of elements is distinguished from another, do not imply associative links within either set of elements but could rely on processes of segregation. Although many physiological theories of grouping are purely associative there are not many methods of isolating associative processes psychophysically. One method has been to measure common resolutions of elements when they are presented under conditions of depth and motion ambiguity. In the present paper a number of the issues surrounding grouping are considered, particularly in relation to associative processes. These include whether grouping is post-constancy and post-completion, grouping with and without emergent features, the role of top-down processing and attention and the explanatory role of general principles such as simplicity and likelihood.

International audience; The present paper provides an analysis of the pointing errors of a patient with unilateral optic ataxia (O.K.) following right hemispheric damage, revealing the type of errors related tothe use of the contralesional hand and/or to the reaching of targets located in the contralesional visual field. In addition, comparison between immediate and delayed pantomime pointing allow testing of whether pointing deficits of this patient are specific to real-time visuo-motor control and, subsequently, whether delay could improve his pointing performance. The results show different patterns in the four hand-field combinations. The following conclusion can be drawn from the results of the delayed condition. In the case of patient O. K., the delay reduced the pointing variability for both hands in the left visual field but not in the right visual field. However, the pointing biases did not improve accordingly. As in healthy subjects, target locations tended to be coded in memory with a bias directed toward the fixation point. These results are discussed and contrasted with respect to those previously obtained in the literature in patients with bilateral optic ataxia.

Because the estimation of thresholds is daily practice in sensory psychophysics, efficient methods must be used to reduce experimental cost and burden. A large number of such methods are available, and each one further has a multitude of variants. All methods presumably provide a threshold estimate that is the stimulus level at which repeated testing would result in a specific percentage of correct responses on a forced-choice task, a percentage that varies across methods and variants thereof. A recent study (García-Pérez, 1998) showed that the most popular method (up-down staircases with fixed step sizes) yields threshold estimates that do not correspond to the presumed percent-correct points. Two modifications of this type of staircase have recently been proposed. In one (Zwislocki and Relkin, 2001), the up-down rule does not require correct responses to occur consecutively. In the other (Kaernbach, 1999), subjects are allowed to respond 'don't know' instead of guessing at random when unsure. Although the statistical basis of either modification were described in general, only a few of their many variants were subjected to evaluation under a limited set of conditions. This paper provides an extensive evaluation of a reasonable number of variants of either modification under a broad set of conditions. The results show that they are generally unfit for threshold estimation because in most cases the percent-correct point that is targeted varies greatly with the relative size of the steps with respect to the spread of the psychometric function. Dependable conditions for the use of these modified staircases are also determined.

Whether contrast adaptation may enhance contrast discrimination is a question that has remained largely unresolved because of conflicting empirical evidence. Greenlee and Heitger (1988), for example, reported that contrast discrimination may be enhanced after contrast adaptation, while Maattanen and Koenderink (1991) did not. This paper aimed to account for the different conclusions reached by these independent researchers by manipulations of key differences that exist between the two studies. It is shown that contrast discrimination may be enhanced after adaptation, but that these effects can vary markedly across subjects and test conditions. Enhancements in contrast discrimination are reported to be significant when adapting and testing at low levels of contrast, but just significant at higher levels of contrast. For high contrast signals; enhancements are shown to be independent of temporal frequency but dependent upon viewing conditions. Under binocular viewing conditions, enhancements in contrast discrimination thresholds are shown to be significantly higher than under monocular viewing conditions. It is suggested that the different conclusions reached by Greenlee and Heitger and by Maattanen and Koenderink may be explained by their respective differences in viewing conditions. The former study used binocular, while the latter study used monocular viewing with an occluding eyepatch.

Relative binocular disparity cannot tell us the absolute 3-D shape of an object, nor its 3-D trajectory if it is moving, unless the visual system has independent access to how far away the object is at any moment. Indeed, as the viewing distance is changed, the same disparate retinal motions will correspond to very different real 3-D trajectories. In this paper we were interested in whether binocular 3-D motion detection is affected by viewing distance. We used a visual search task in which the observer is asked to detect a target dot, moving in 3-D, amidst 3-D stationary distractor dots. We found that distance does not affect detection performance. Motion-in-depth is consistently harder to detect than the equivalent lateral motion, for all viewing distances. For a constant retinal motion with both lateral and motion-in-depth components, detection performance is constant despite variations in viewing distance that produce large changes in the direction of the 3-D trajectory. We conclude that binocular 3-D motion detection relies on retinal, not absolute visual signals.

This paper is organised approximately into two halves. In the first half, I review evidence about the structure of the visual system, and I use that evidence to frame what I think are widely held but often implicit ideas about how it works, namely that vision is principally analysis of retinal input. These ideas have been strongly influenced by engineering approaches; form a default view of the visual system that suffuses all the language used to describe it (at least in visual neuroscience); and are to some extent supported by the structural evidence. In the second half, I explore some inconvenient facts from neuroanatomy and neurophysiology which are quite uncomfortable for the traditional view. I then set out a contrary view of how structure and function are linked in the visual system, which is a neurobiological variety of the quite developed view in psychophysics that vision is better understood as knowledge-rich inference. Finally, I explore some of the ramifications of this view for neurophysiological understanding of how the visual system might operate during normal vision.

Most current models of human and animal vision assume that the processes of vision involve 2D (or even 3D) internal representations of the external world - an iconic representation. Within these models, recognition involves some form of lineal, areal or volumetric comparison of these internal representations (either learned or innate) with current sensory inputs. However, this view has recently come under criticism. In this paper, a neural model of vision is explored in which this iconic world view is replaced by a temporally ordered trace of (essentially) local features. The model employs hierarchical, recurrently linked, self-organizing topological maps.

Forced-choice staircases with fixed step sizes are used in a variety of formats whose relative merits have never been studied. This paper presents a comparative study aimed at determining their optimal format. Factors included in the study were the up/down rule, the length (number of reversals), and the size of the steps. The study also addressed the issue of whether a protocol involving three staircases running for N reversals each (with a subsequent average of the estimates provided by each individual staircase) has better statistical properties than an alternative protocol involving a single staircase running for 3N reversals. In all cases the size of a step up was different from that of a step down, in the appropriate ratio determined by García-Pérez (Vision Research, 1998, 38, 1861 - 1881). The results of a simulation study indicate that a) there are no conditions in which the 1-down/1-up rule is advisable; b) different combinations of up/down rule and number of reversals appear equivalent in terms of precision and cost: c) using a single long staircase with 3N reversals is more efficient than running three staircases with N reversals each: d) to avoid bias and attain sufficient accuracy, threshold estimates should be based on at least 30 reversals: and e) to avoid excessive cost and imprecision, the size of the step up should be between 2/3 and 3/3 the (known or presumed) spread of the psychometric function. An empirical study with human subjects confirmed the major characteristics revealed by the simulations.

This paper introduces a new approach to assess visual representations underlying the recognition of objects. Human performance is modeled by CLARET, a machine learning and matching system, based on inductive logic programming and graph matching principles. The model is applied to data of a learning experiment addressing the role of prior experience in the ontogenesis of mental object representations. Prior experience was varied in terms of sensory modality, i.e. visual versus haptic versus visuohaptic. The analysis revealed distinct differences between the representational formats used by subjects with haptic versus those with no prior object experience. These differences suggest that prior haptic exploration stimulates the evolution of object representations which are characterized by an increased differentiation between attribute values and a pronounced structural encoding.

The pattern of lateral interactions in the primary visual cortex, which has emerged from recent studies, conforms to the grouping rules of similarity, proximity, smoothness and closure. The goal of this paper is to understand the perceptual salience of oriented elements that are specifically organized to form a smooth contour. An overview of recent studies, in combination with new experimental results, is presented here to emphasis the idea that visual responses depend on input from both the center and the surround of the classical receptive field (CRF). It is assumed that normal lateral interactions produce a neuronal network that is formed by two antagonistic mechanisms: (i) excitation, that is spatially organized along the optimal orientation (collinear), and is predominant near the contrast threshold of the neuron, and (ii) inhibition, that is less selective and is distributed diffusely around the cell's response field. Thus, the inputs from the CRF and the anisotropic surround are summated non-linearly. The specificity of the facilitation and suppression along the collinear direction suggests the existence of second-order elongated collinear filters, which may increase the response similarity between neurons responding to elongated stimulus, thus may enhance the perceptual salience of anisotropic configurations such as contours. This causal connection is particularly evident in amblyopes, where abnormal development of the network results in the abnormal perception of contours.

Psychophysical studies show that the duration of visual persistence increases with spatial frequency of gratings. Previous theories ascribe this finding to differences between the spatial and temporal properties of sustained and transient pathways. This paper proposes an alternative account that explains persistence as a side-effect of excitatory feedback in neural circuits for contour extraction. Mechanisms to break excitatory feedback include inhibitory reset signals at stimulus offset. Simulations demonstrate how gratings with lower spatial frequency generate stronger inhibitory reset signals, thereby resulting in shorter persistence for lower spatial frequencies. Additional simulations account for interactions of spatial frequency with stimulus duration, effects of adaptation, and properties of residual traces, as opposed to visual persistence.

Visual neurons in the primary visual cortex 'look' at the retinal image through a four-dimensional array of spatial receptive fields (filter-elements): two spatial dimensions and, at each spatial location, two Fourier dimensions of spatial frequency and orientation. In general, visual objects activate filter-elements along each of these dimensions, suggesting a need for some kind of linking mechanism that determines whether two or more filter-elements are responding to the same or different contours or objects. In the spatial domain, a (spatial) association field between filter-elements, arranged to form first-order curves, has been inferred as a flexible method by which different parts of extended (luminance) contours become associated (Field et al., 1993). Linking has also been explored between filters selective for different regions in Fourier space (e.g. Georgeson and Meese, 1997). Perceived structure of stationary plaids suggests that spatial filtering is adaptive: synthetic filters can be created by the linear summation of basis-filters across orientation or spatial frequency in a stimulus-dependent way. For example, a plaid with a pair of sine-wave components at +/-45 deg looks like a blurred checkerboard; a structure that can be understood if features are derived after linear summation of spatial filters at different orientations. However, the addition of an oblique third-harmonic component causes the plaid to perceptually segment into overlapping oblique contours. This result can be understood if filters are summed across spatial frequency, but, in this case, treated independently across orientation. In the present paper, the architecture of an association field is proposed to permit linking and segmentation of filter-elements across spatial frequency and orientation. Three types of link are proposed: (1) A chain of constructive links around sites of common spatial frequency but different orientation, to promote binding of filters across orientation; (2) Constructive links between sites with common orientation but different spatial frequency, to promote binding of filters across spatial frequency; (3) Long-range links between sites of common spatial frequency but different orientation, whose activation and role are determined by activity in a higher spatial frequency band. A model employing the proposed network of links is consistent with at least six previously reported effects on the perception of briefly presented stationary plaids.

Rogers and Graham (1979) developed a system to show that head-movement-contingent motion parallax produces monocular depth perception in random dot patterns. Their display system comprised an oscilloscope driven by function generators or a special graphics board that triggered the X and Y deflection of the raster scan signal. Replication of this system required costly hardware that is no longer on the market. In this paper the Rogers-Graham method is reproduced with an Intel processor based IBM PC compatible machine with no additional hardware cost. An adapted joystick sampled through the standard game-port can serve as a provisional head-movement sensor. Monitor resolution for displaying motion is effectively enhanced 16 times by the use of anti-aliasing, enabling the display of thousands of random dots in real-time with a refresh rate of 60 Hz or above. A color monitor enables the use of the anaglyph method, thus combining stereoscopic and monocular parallax on a single display without the loss of speed. The power of this system is demonstrated by a psychophysical measurement in which subjects nulled head-movement-contingent illusory parallax, evoked by a static stereogram, with real parallax. The amount of real parallax required to null the illusory stereoscopic parallax monotonically increased with disparity.

This paper presents three experiments investigating the claim that the visual system utilizes a primitive indexing mechanism (sometimes called FINSTS; Pylyshyn, 1989) to make non-contiguous features directly accessible for further visual processing. This claim is investigated using a variant of the conjunction search task in which subjects search among a subset of the items in a conjunction search display for targets defined by a conjunction of colour and orientation. The members of the subset were identified by virtue of the late onset of the objects' place-holders. The cued subset was manipulated to include either homogeneous distractors or mixed distractors. Observers were able to select a subset of three items from among fifteen for further processing (Experiment 1); furthermore, a reaction time advantage for homogeneous subsets over mixed subsets was observed, indicating that more than one of the subset is selected for further specialized processing. The homogeneous subset advantage held for subsets of two to five items (Experiment 2), and the time required to process the cued subset did not increase with increased dispersion of the items (Experiment 3). These results support the basic claim of the indexing theory: the claim that multiple visual indexes are used in selecting objects for visual processing.

It has been shown that for every model of detection (whether single- or multi-channel, with linear or non-linear channels, and whatever decision rule), provided that it predicts a convex set of subthreshold stimuli, there is a psychophysically equivalent peak detector made up of a collection of linear analysers followed by a maximum-output decision rule. In this paper, the equivalent peak detector representations of some widely accepted detection models are calculated. The calculations rest on a general technique for deriving, from a given model, a formula which specifies the analyser most sensitive to any given stimulus.

Abstracet-Many researchers have proposed that objects are perceived as structural descriptions, which specify the configuration of an object's features (or parts) in terms of their categorical relations to one another. Others have proposed that objects arc perceived as views, which specify the configuration of an object's features in terms of their coordinates, in particular 2D views. This paper presents five experiments testing these competing accounts of the perception of the configuration of an object's features. Subjects learned to recognize a set of target objects and were tested for their ability to distinguish them from various distractors that differed either in their categorical relations or their coordinates. Subjects were consistently more likely to confuse both 2D and 3D objects that were similar in their parts' relations to each other than to confuse objects similar in their parts' coordinates (in any reference frame). This effect persisted when subjects were allowed to view the objects as long as they wished and when they were explicitly trained to distinguish them from the distractors. These findings suggest that we perceive an object's features in terms of their categorical relations to one another. A preliminary model of the findings is presented.

Under appropriate experimental conditions, the threshold intensity of a visual stimulus varies as the square-root of the background illuminance. This square-root law has been observed in both psychophysical threshold experiments and in measurements of the thresholds of individual ganglion cells. A signal detection theory developed in the 1940s by H. L. de Vries and A. Rose, and since elaborated by H. B. Barlow and others, explains the square-root law on the basis of 'noise' due to fluctuations in the number of photon absorptions per unit area and unit time at the cornea. An alternative account of the square-root law - and also other threshold-vs-intensity slopes - is founded on the assumption of physiological gain control (W. A. H. Rushton, Proc. Roy. Soc. (London) B 162, 20-46, 1965; W. S. Geisler, J. Physiol. (London) 312, 165-179, 1979). In this paper, a neural model of light adaptation and gain control is described that shows how these two accounts of the square-root law can be reconciled by a stochastic gain control mechanism whose gain depends on the photon fluctuation level. The process by which spikes are generated in a ganglion cell is modeled in terms of a stochastic integrate-and-fire mechanism; this model is used to quantitatively fit toad retinal ganglion cell threshold data. A psychophysical model is then outlined showing how a statistical observer could analyze the ganglion cell spike trains generated by 'signal' and 'noise' trials in order to statistically discriminate the two conditions. The model is also shown to account for some dynamic aspects of ganglion cell responses, including ON- and OFF-responses. The neural light adaptation model predicts that - under the proper conditions - brightness matching judgments will also be subject to a square-root law. Experimental tests of the model under superthreshold conditions are proposed.

A translation of von Krie's (1882) paper "On the measurement of intensive magnitudes and on the so-called psychophysical law" is accompanied by a commentary. Von Kries claims that intensive magnitudes are not measurable in themselves, because the establishment of an equivalence between different steps in a scale of intensity does not make any sense without further clarification. Where intensive magnitudes are determined in a domain of the natural sciences, he claims this is only a matter of counting, and of the measurement of temporal and spatial magnitudes. Every measurement of intensity should then be reduced to these operations by explicit conventions. Likewise, we can only speak of the measurement of sensations once we have established an arbitrary convention that determines what we will consider equal. The debate whether sensation varies with the logarithm of stimulus intensity or in direct proportion to stimulus intensity, is then not a difference over matters of fact. Instead it is an empty dispute over words that is rooted in misunderstanding.

A wealth of detection data can be accounted for by a spatial-vision model including a finite number of space-variant, spatial-frequency and orientation-selective channels of varying gains coupled with a detection rule involving probability summation over space and among channels. This paper shows that the detection of large-area, foveally fixated sine-wave gratings can be understood as if it occurred merely as a result of the activity of the subset of channels whose orientation matches that of the gratings, and operating under a peak-detection rule. This simplification makes it possible to show the theoretical relationship between the large-area sine-wave contrast sensitivity function, the channel gain function, and the channel modulation-transfer functions. It is also shown that the human visual system must have many more channels than are normally assumed in spatial-vision models, for otherwise the contrast sensitivity function would show significant bumps. An unlimited-channel model with a given mathematical form for the channels' modulation transfer functions is used to derive a mathematical expression for the (foveal) channel gain function from the large-area psychophysical sine-wave contrast sensitivity function.

This paper reviews empirical evidence for the detection of visual symmetries and explanatory theories and models of symmetry detection. First, mirror symmetry is compared to other types of symmetry. The idea that symmetry detection is preattentive is then discussed and other roles that attention might play in symmetry detection are considered. The major part of the article consists of a critical examination of the extensive literature about the effects on symmetry detection of several major factors such as the orientation of the symmetry axis, the location of the stimulus in the visual field, grouping, and perturbations. Constraints on plausible models of symmetry detection are derived from this rich database and several proposals are evaluated against it. As a result of bringing this research together, open questions and remaining gaps to be filled by future research are identified. ispartof: Spatial Vision vol:9 issue:1 pages:9-32 ispartof: location:Netherlands status: published

This paper describes an investigation of eye movements made by eighteen observers with normal spatial vision in response to eleven images of natural scenes each of which was presented in three versions, unfiltered filtered by low-pass and filtered by high-pass spatial frequency filters. The ability of observers to identify the different images was determined after each set of measurements. An index of similarity, calculated in terms of the sum of the squares of the distances between each fixation point and its nearest neighbour in the other set, was developed in order to provide a basis for the comparison of two sets of eye movements, each made by a different observer or both made on different occasions by the same observer. Application of this index shows that brief (1.5-s) presentation, there exists a high degree of similarity between fixations made by different observers in response to the same image. For longer (3-s) periods, however, the similarity for inter-observer comparisons is reduced. The data demonstrate that filtering of the image has little effect on the pattern of fixations made to a given image, but the duration of fixations is greatest for low-pass filtered images and least for unfiltered images, whereas the amplitude of saccades is greatest for unfiltered images and least for low-pass filtered images. It is proposed that for brief presentations, eye movements made during examination of an unfamiliar image are performed automatically in response to the spatial features of the image.

Models of the spatial response of human vision are important for applied work, but the available contrast sensitivity function (CSF) data vary widely due to the diverse spatiotemporal stimuli used over the years. To assist selection, this paper: (1) reports measurements of the effects on the CSF of varying the spatial and temporal windows of grating patches; (2) demonstrates that the widely discrepant CSFs from previous studies can be accounted for by using these results; and (3) discusses simple criteria for choosing CSFs for practical applications. CSFs were measured for several combinations of spatial and temporal waveforms, using the same subjects under otherwise identical conditions. The CSF was measured over the range of 0.5-10 c/deg using Gabor-type patches of 1.0-, 0.5-, 0.25-, and 0.125-octave spatial bandwidths using both abrupt and gradual temporal presentations. The results were compared with the CSF obtained with a fixed aperture (4 deg x 4 deg) grating pattern. Increasing the number of cycles resulted in increased sensitivity at intermediate frequencies, changing the CSF to a narrower bandpass shape. For each patch bandwidth, the gradual presentation CSF had a narrower spatial pass band than with the abrupt presentation. The relevance of the large differences in the CSFs obtained with different stimuli to our understanding of visual performance is discussed.

In this paper we define the content of information in an image and show how it can be computed by taking into account different levels of resolution, in the framework of information theory and the thermodynamics of irreversible transformations. The results thus obtained will eventually be exploited to derive a mechanism for active exploration of visual space suitable to perform a dynamic coupling between the agent and its environment.

Recent discussion regarding whether the noise that limits 2AFC discrimination performance is fixed or variable has focused either on describing experimental methods that presumably dissociate the effects of response mean and variance or on reanalyzing a published data set with the aim of determining how to solve the question through goodness-of-fit statistics. This paper illustrates that the question cannot be solved by fitting models to data and assessing goodness-of-fit because data on detection and discrimination performance can be indistinguishably fitted by models that assume either type of noise when each is coupled with a convenient form for the transducer function. Thus, success or failure at fitting a transducer model merely illustrates the capability (or lack thereof) of some particular combination of transducer function and variance function to account for the data, but it cannot disclose the nature of the noise. We also comment on some of the issues that have been raised in recent exchange on the topic, namely, the existence of additional constraints for the models, the presence of asymmetric asymptotes, the likelihood of history-dependent noise, and the potential of certain experimental methods to dissociate the effects of response mean and variance.

Cavanagh and Mather (1989) reviewed literature concerning the possible distinction between short- and long-range processes in motion perception and concluded that the distinction cannot be supported. Instead, they proposed that motion perception be considered on the basis of detectors for first-order (luminance, color) and second-order (first-order motion, texture, stereo) stimulus attributes. They supported their position with studies of motion based on second-order stimuli. The present paper contends that when experiments permitting the investigation of both processes in the same display are included and when criteria are examined in their totality rather than one-by-one, the original short-range/long-range distinction can be retained. Furthermore, it is argued that the first-order/second-order distinction does not represent a theoretical advancement and that studies of second-order motion can be interpreted in terms of the older distinction. It is concluded that the short-range/long-range distinction is useful and should not be abandoned.

This paper calls attention to research showing that binocular disparity, which is an effective cue to depth, plays a secondary role, at best, in the perception of 3D shape. This claim has implications both for how shape should be studied and how this unique perceptual property should be modeled. These issues are discussed from a historical perspective, which shows how the failure to appreciate the importance of the Gestalt grouping principle called 'Figure-Ground Organization' led to many unfruitful efforts. It also calls attention to how this situation can be remedied.

Representational art tries to capture a 3D world on a 2D surface, and artists often discuss this in relation to the projected image on window panes and mirrors. But are 2D projections on transparent surfaces useful to learn about projections in general? Most people are unaware of the 2D projected size of objects on the surface of mirrors. They also incorrectly expect that these projections always get smaller with distance of the target object from the mirror, and do not change with distance of the observer (when the target is stationary). In this paper we extend this result about surfaces of mirrors to surfaces of windows, and we confirm that the errors that people make are not specific to Western culture by replicating the study in China. In contrast to their errors about projections, people are more accurate at predicting how field of view will vary depending on distance of the observer from a mirror or window. To explain how this pattern of (false) beliefs can stem from experience we argue that people do not perceive projections on transparent surfaces.

The ability to locate an object in the visual field is a collaboration of at least three intermingled processes: scanning multiple locations, recognizing the object sought (the target), and ending the search in cases when the target is not found. In this paper, we focus on the termination rule. Using distribution analyses, it is possible to assess the probability of termination conditional on the number of locations examined. The results show that on some trials without target, the participants carried out more comparisons than there are objects in the display; in other conditions, they carried out fewer comparisons than objects. Because there were very few errors, the premature stops were not pure guesses. We present models to account for these findings. The distributions of terminations help determine the slopes of the functions relating response time to set size.

This paper is concerned with the information used in open-loop pointing to visually perceived targets. Stereoscopic stimuli were used to produce illusory relative egocentric distances, which were inconsistent with the angles of vergence required to fuse the targets. One of the stimuli was a rectangle slanted around a vertical axis. Four participants in Experiment 1 reported its slant and pointed to its edges. The slant was hugely underestimated (condition A) unless the rectangle was flanked by other surfaces (condition B). The relative depth of a pair of dots placed in front of the rectangle was also misperceived due to depth-contrast effect. The critical finding is that pointing responses were not based on vergence but were consistent with depth estimates, both for the rectangle and for the dots. Experiment 2 revealed the conditions necessary for pointing to be consistent with perceived relative position. The different target distances were either randomised allowing inter-trial comparisons, or presented only one per session to prevent them. Pointing was similar to estimates only in the randomised condition showing the significance of inter-trial comparisons. It is proposed that participants used the remembered motor command and kinesthetic sensations of a previous movement as a reference, attempting to make the difference between successive movements the same as a visually perceived depth difference between successive targets.

This paper reports a comparison between two tasks of visual search. Two observers carried out, in separate blocks, a saccade-to-target task and a manual-target-detection task. The displays, which were identical for the two tasks, consisted of a ring of eight equally spaced Gabor patches. The target could be defined by a difference from the distractors along four possible dimensions: orientation, spatial frequency, contrast or size. These four dimensions were used as variables in separate experiments. In each experiment, performance was measured over an extensive range of values of the particular dimension. Thresholds were thus obtained for the saccade and the manual response tasks. The nature of the response was found to modify the relative visual sensitivity. For orientation differences, manual response performance was better than saccade-to-target performance. The reverse was true for spatial frequency and contrast differences, where saccade-to-target performance was better than manual response performance. We conclude that saccade-selection in a search task draws on different visual information from that used for manual responding in the equivalent task. The two tasks thus differ in more than the different response systems used: the results suggest the action of different underlying neural visual mechanisms as well as different neural motor mechanisms.

As opposed to dealing with the geometry of objects in the 3D world, this paper considers the geometry of the visual input itself, i.e. the geometry of the spatio-temporal hypersurface defined by image intensity as a function of two spatial coordinates and time. The results show how the Riemann curvature tensor of this hypersurface represents speed and direction of motion, and thereby allows to predict global motion percepts and properties of MT neurons. It is argued that important aspects of early and middle level visual coding may be understood as resulting from basic geometric processing of the spatio-temporal visual input. Finally, applications show that the approach can improve the computation of motion.

The problem of learning is arguably at the very core of the problem of intelligence, both biological and artificial. In this paper we sketch some of our work over the last ten years in the area of supervised learning, focusing on three interlinked directions of research: theory, engineering applications (that is, making intelligent software) and neuroscience (that is, understanding the brain's mechanisms of learning).

The goal of the present paper is to propose a solution to the 'saliency problem' which has been raised in regard to Rakover and Cahlon's (1989) Catch model for identifying a previously seen target face (Ft). In contrast to real life situations, the Catch model assigned the same weight to different facial dimensions and values. Mathematical proofs, reanalyses of the results of three experiments reported in Rakover and Cahlon, and the analysis of the results of a new experiment show that this proposal expands and improves the Catch model.

The aim of this paper is to demonstrate the presence of a similar visual field effect on both spatial orientation and egocentric localization. A first experiment explored an orientation task (Visually Perceived Vertical or VPV determination) and compared the effects of a frame inclined either in the midfrontal plane (in this condition subjects assessed roll VPV) or in the median plane (subjects assessed pitch VPV) or in both combined planes (subjects assessed both roll and pitch VPV). A second experiment compared the frame effects specified above to the frame effect observed in an egocentric localization task (Visually Perceived Eye Level judgment) performed with a frame slanted in the median plane. The results showed that angular frame variations from -15 to +15 deg result in the same psychometric function for both orientation and localization tasks. In each experiment, correlations showed that individual differences occur in relation to an overall sensitivity to the visual field. Individual sensitivity may be accounted for by a ratio of visual to graviceptive information which remains constant whatever the perception plane (midfrontal or median plane) and whatever the task (spatial orientation or localization).

Stereoscopic vision is extremely precise in detecting minute differences between adjacent depth planes, but quite imprecise in estimating absolute depth. In this paper, we address the issue of the spatial acuity (and not the stereo acuity) of stereopsis. Static RDS (random dot stereograms) stimuli were used to find the spatial grain in which human stereoscopic vision operates. Using psychophysical experiments it was found that foveally, stimuli smaller than 8' cannot be accurately perceived. For other eccentricities, it was found that this threshold is inversely proportional to the Cortical Magnification factor. We interpret this spatial size limit, which is an order of magnitude larger than visual spatial acuity, as an indication that stereopsis is an area based comparison rather than a point process, and discuss the relations between the cortical 'patch' size that corresponds to this 8' limit and Ocular Dominance Columns.

Inexpensive circuit boards have appeared on the market which transform a normal micro-computer's disk drive into a video disk capable of playing extended video sequences in real time. This technology enables the performance of experiments which were previously impossible, or at least prohibitively expensive. The new technology achieves this capability using special-purpose hardware to compress and decompress individual video frames, enabling a video stream to be transferred over relatively low-bandwidth disk interfaces. This paper will describe the use of such devices for visual psychophysics and present the technical issues that must be considered when evaluating individual products.