Vision Disparity

Abstract

Vision disparity, the binocular depth cue that the Medical Subject Headings vocabulary files under depth perception, is the small difference between the two eyes' retinal images that arises because the eyes view the world from slightly separated vantage points. The brain reads this geometric signal to recover depth, a process called stereopsis. This article treats disparity as the input to a computation rather than as a percept in itself: how the horizontal offset between corresponding image points is defined and measured, why the visual system must first solve a correspondence problem to know which points in the two images belong together, and how disparity-selective neurons from primary visual cortex onward encode the cue. It covers stereoacuity, the random-dot stereogram, the disparity-energy model, and the comparative reach of stereopsis across the animal kingdom.

Keywords: binocular disparity, stereopsis, correspondence problem

What Vision Disparity Is

Vision disparity is the difference between the images the two eyes receive of the same scene. Because the eyes are set roughly 6.5 centimetres apart, each looks out from a slightly different position, and any object that is nearer or farther than the point of fixation projects to non-corresponding places on the two retinas. That geometric offset, measured as an angle, is the binocular disparity. It is not itself a sense of depth; it is the raw signal from which the brain computes depth, and it is best understood as the input to a visual computation rather than as something the observer consciously sees.

The distinction between the cue and the percept it feeds is the organizing idea of this article, and it is also why the Medical Subject Headings vocabulary treats vision disparity as a classification beneath depth perception rather than as a synonym for it. Charles Wheatstone made the cue visible as a cause in 1838 when he built the stereoscope and showed that presenting each eye with a slightly different flat drawing, differing only in the horizontal positions of their features, produced a vivid impression of solid depth where none existed in either image alone (Wheatstone, 1838). Disparity, he established, is sufficient for depth: the two-dimensional images carry no depth of their own, yet their difference does.

Disparity comes in several forms, and keeping them apart matters because they are read by different mechanisms and support different judgments. Table 1 lays out the principal distinctions: the sign of the offset, which tells near from far; whether the disparity is measured against the fixation point or between two objects; and the direction of the offset, only the horizontal component of which the geometry of the two eyes makes informative about depth.

Measuring Disparity: Stereoacuity and Geometry

Disparity is an angle, and the visual system is extraordinarily sensitive to it. Under good conditions a human observer can detect a disparity of only a few arc seconds, a resolution finer than the spacing of the photoreceptors and so classed as a hyperacuity, a discrimination better than the sampling grain of the retina would naively allow (Westheimer, 1979). This threshold, the smallest disparity that supports a reliable depth judgment, is stereoacuity, and it is the standard psychophysical handle on the cue.

The geometry that converts a depth difference into a disparity is simple and has a consequence worth stating plainly. For a point a small distance in front of or behind the fixation point, the disparity it generates is, to a close approximation, proportional to the interocular separation and to the depth difference, and inversely proportional to the square of the viewing distance. The inverse-square term is why stereopsis is a near-space sense: the same physical depth step that produces an easily seen disparity on a desktop produces a vanishingly small one across a field. The range over which binocular depth discrimination operates, and the way it falls off with distance, was mapped systematically by Blakemore (Blakemore, 1970). The DisparityGeometryDemo lets the reader move an object in depth and watch the two retinal projections, and the resulting disparity, change.

Two further facts about measurement matter. First, only disparities within a limited range around the horopter, the locus of points that project to corresponding retinal locations, are fused into a single percept; this zone is Panum's fusional area, and disparities beyond it are seen as double images even though they can still drive a sense of depth. Second, the cue the visual system privileges for fine depth is relative disparity, the difference between two objects' absolute disparities, because relative disparity is invariant to small errors in eye position that corrupt the absolute signal. The result is a system optimized for judging the depth between nearby surfaces rather than the absolute distance of any one of them.

The Correspondence Problem

Before the visual system can measure a disparity, it faces a prior and surprisingly hard problem: it must decide which feature in the left image corresponds to which feature in the right. Only once two image points are known to be projections of the same world point does the offset between them count as a disparity. When the scene is full of similar features, many false pairings are geometrically possible, and the great majority of them are wrong. This is the correspondence problem, and it is the computational heart of stereopsis.

Béla Julesz proved that the brain solves correspondence on the raw image, before any object is recognized, with an elegant stimulus: the random-dot stereogram. Each eye is shown a field of random dots that is meaningless alone; a central region is shifted horizontally in one eye's image, introducing a disparity but no monocular contour, shape, or familiarity cue. Fused, the dots resolve into a surface floating in depth, demonstrating that disparity alone, extracted from an image with no recognizable form, is sufficient to see depth (Julesz, 1964). The RandomDotStereogramDemo generates such a stereogram deterministically and lets the reader set the disparity of the hidden figure.

The random-dot stereogram made the correspondence problem concrete, and David Marr and Tomaso Poggio gave it a computational theory. They argued that the brain narrows the field of possible matches using constraints drawn from the physics of surfaces: a given point in one image has a single depth, so it should match at most one point in the other (uniqueness), and because matter is cohesive, disparity should vary smoothly almost everywhere (continuity). A cooperative algorithm that enforces these constraints converges on the one globally consistent set of matches (Marr & Poggio, 1979). Not every point can be matched, and the visual system exploits this too: points visible to one eye but occluded from the other carry their own information about depth and edges, the phenomenon of da Vinci stereopsis (Nakayama & Shimojo, 1990). A modern proposal reframes the matching problem in neural terms, suggesting that the brain devotes a population of “what not” detectors to actively signalling and suppressing the false matches, so that the surviving true matches define the surface (Goncalves & Welchman, 2017).

Neural Encoding of Disparity

Where does disparity become a neural signal? The first disparity-selective neurons were found by Horace Barlow, Colin Blakemore, and Jack Pettigrew in the cat's visual cortex: cells that fired best when a stimulus fell on the two retinas at a particular relative offset, tuned to a preferred disparity and so capable of acting as a depth filter (Barlow, Blakemore, & Pettigrew, 1967). Gian Poggio then showed that the awake, behaving monkey's striate and prestriate cortex contains neurons tuned to disparity, including cells selective for the near, the far, and the plane of fixation, establishing a cortical substrate for stereopsis in the primate (Poggio & Fischer, 1977).

How a neuron computes disparity is captured by the disparity-energy model. Izumi Ohzawa, Gregory DeAngelis, and Ralph Freeman recorded binocular simple and complex cells whose responses are predicted by combining the two eyes' inputs through receptive fields that differ between the eyes, yielding a cell tuned to a preferred disparity largely independent of the stimulus's exact position (Ohzawa, DeAngelis, & Freeman, 1990). The two eyes' receptive fields can differ in two ways, a shift in position or a shift in the phase of their internal structure, and real cortical neurons use a mixture of both, a detail that constrains how the disparity map is built (Anzai, Ohzawa, & Freeman, 1999). The DisparityTuningDemo plots the tuning curve of a model energy neuron and lets the reader shift its preferred disparity between near-tuned, far-tuned, and tuned-excitatory types.

Encoding the cue is not the same as perceiving depth, and the two can be prised apart. Bruce Cumming and Andrew Parker showed that neurons in primary visual cortex respond to binocular disparity even in anticorrelated stereograms, where one eye's dots are contrast-reversed and no coherent depth is seen, so V1 signals the raw disparity before the stage that decides perceived depth (Cumming & Parker, 1997). The causal link to perception appears later in the visual hierarchy: electrically microstimulating disparity-tuned columns in area MT biases a monkey's depth judgments toward the stimulated cells' preferred disparity, tying an extrastriate area to the percept itself (DeAngelis, Cumming, & Newsome, 1998). The architecture that turns a V1 disparity signal into a depth judgment, through specialized pathways and cortical areas, is the subject of Parker's synthesis of binocular depth perception in the cerebral cortex (Parker, 2007). Figure 1 traces the path from the geometric cue to the perceptual decision.

Figure 1

Worked Example: From Depth Step to Disparity

The inverse-square law is worth working through, because it explains the character of stereopsis better than any verbal claim. Take the interocular distance as I = 65 mm and fix the eyes on a point at D = 1.0 m. A second object sits a small distance δ = 5 cm beyond the fixation point. The relative disparity it generates is, to close approximation, η ≈ I × δ / (D × (D + δ)), which evaluates to 0.065 × 0.05 / (1.0 × 1.05) = 0.00310 radians. Converting to arc seconds, multiply by 206,265 to get about 638 arc seconds, roughly 10.6 arc minutes, a large and easily seen disparity.

Now run the law the other way to see the price of distance. Stereoacuity near threshold is about 5 arc seconds; the smallest depth step a person can resolve at distance D is δ_min ≈ η_threshold × D² / I. At half a metre this is 0.09 mm; at one metre it is 0.37 mm; at two metres it has grown to 1.49 mm. Doubling the viewing distance, from 0.5 m to 1.0 m and again to 2.0 m, degrades the finest resolvable depth step by a factor of four each time, the signature of the D² term. The arithmetic makes the point the geometry promised: binocular disparity is a precision depth sense for the space within arm's reach and a progressively coarser one beyond it, which is exactly the range over which grasping, tool use, and manipulation demand fine depth.

Discussion

Vision disparity is the clearest case in perception of a cue that is not a percept. The eyes deliver a geometric difference; the brain performs a computation on it; depth is the output. Separating these three things, the signal, the computation, and the perceptual result, is what lets the topic be studied with precision, and it is why the Medical Subject Headings vocabulary classifies vision disparity beneath depth perception rather than equating them. The disparity is real and measurable in the images; whether it yields a depth percept depends on stages of processing that can be probed, lesioned, and stimulated independently of the cue itself.

The field's durable achievement is to have closed the loop from geometry to neuron to perception. Wheatstone showed the cue was sufficient for depth; Julesz showed the brain extracts it before recognizing any object; Marr and Poggio specified the computation; Barlow and Poggio found the neurons; Ohzawa and colleagues gave the neuron a model; and Cumming, Parker, DeAngelis, and Newsome dissected which stage carries the raw cue and which carries the percept. Few topics in cognitive psychology can trace such a complete chain from physics to phenomenology.

What remains open is instructive. The correspondence problem is solved by the brain with a speed and robustness that still outstrips most machine-vision systems, and the neural implementation of the global consistency that Marr and Poggio's theory requires is not fully understood. The role of suppressive, false-match-rejecting signals, the contribution of vertical disparities to calibrating viewing distance, and the way disparity is combined with the other depth cues into a single estimate are all active questions rather than settled ones.

Current Directions

A major current thread is the read-out problem: given that V1 encodes disparity but does not by itself produce the percept, how do later areas combine, denoise, and interpret the population signal to yield a stable three-dimensional scene? Human neuroimaging and psychophysics have begun to map how the brain integrates disparity with other cues and resolves it into a depth estimate (Welchman, 2016), and the neural architectures that carry stereo information through parallel cortical pathways are being characterized in detail (Parker, Smith, & Krug, 2016). The proposal that dedicated “what not” detectors actively suppress false matches offers a concrete neural mechanism for the correspondence computation that had been treated algorithmically (Goncalves & Welchman, 2017).

A second thread is comparative and evolutionary. Stereopsis is not a primate peculiarity: it has been demonstrated across a striking range of animals, from birds of prey to the praying mantis, and comparing how different nervous systems solve the same disparity computation separates the general principles from the mammalian implementation (Nityananda & Read, 2017). Surveying binocular vision across the animal kingdom shows that disparity-based depth has evolved repeatedly, with different front-end optics and neural circuits converging on the same geometric cue, which both constrains theories of the computation and opens simpler model systems for studying it (Read, 2021).

Common Misconceptions

Vision disparity is the same thing as depth perception.

Disparity is one binocular cue to depth, not depth perception itself. The brain also uses motion parallax, occlusion, shading, and perspective, and it computes a depth percept from disparity rather than reading it off directly, which is why MeSH files vision disparity beneath depth perception (Parker, 2007).

Seeing depth from disparity requires recognizing objects first.

The random-dot stereogram disproves this: observers see a surface in depth from disparity alone, with no recognizable form, edge, or familiarity available in either monocular image, so the brain extracts disparity before it recognizes anything (Julesz, 1964).

A neuron that responds to disparity is perceiving depth.

Primary visual cortex neurons respond to disparity even in anticorrelated stereograms that yield no depth percept, showing that encoding the cue is dissociable from perceiving depth; the percept depends on later stages such as area MT (Cumming & Parker, 1997).

Stereopsis works equally well at any distance.

The disparity a fixed depth step produces falls off with the square of viewing distance, so stereoacuity is exquisite within arm's reach and coarse across a room; binocular disparity is fundamentally a near-space depth sense (Blakemore, 1970).

Glossary

Absolute disparity.

The retinal offset of a single point relative to the current point of fixation; signals that point's depth relative to where the eyes converge.

Anticorrelated stereogram.

A stereogram in which one eye's pattern is contrast-reversed; it drives disparity-tuned neurons in V1 yet produces no coherent depth percept, dissociating the cue from perception.

Binocular disparity.

The angular difference between the two eyes' retinal images of a point, arising from the eyes' horizontal separation; the geometric cue the brain reads to recover depth.

Correspondence problem.

The computational problem of deciding which point in the left image is the same world point as a given point in the right, which must be solved before a disparity can be measured.

Crossed disparity.

The disparity sign produced by an object nearer than the fixation plane; signals that the object is closer than the point of convergence.

Depth perception.

The perception of the three-dimensional layout of a scene from many cues; the broader MeSH category under which vision disparity is filed.

Disparity-energy model.

The standard model of how a binocular neuron computes disparity, combining the two eyes' receptive-field inputs so the cell is tuned to a preferred disparity largely independent of stimulus position.

Horopter.

The locus of points in space that project to corresponding locations on the two retinas and so have zero disparity for a given fixation.

Hyperacuity.

A discrimination finer than the spacing of the photoreceptors; stereoacuity is a hyperacuity, resolving disparities of only a few arc seconds.

Panum's fusional area.

The range of disparities around the horopter within which the two eyes' images are fused into a single percept rather than seen double.

Random-dot stereogram.

A pair of random-dot fields, meaningless alone, in which a disparity-shifted region yields a surface in depth when fused; Julesz's proof that disparity alone is sufficient for depth.

Relative disparity.

The difference in absolute disparity between two points; the signal the visual system privileges for fine depth because it is invariant to small errors in eye position.

Stereoacuity.

The smallest disparity that supports a reliable depth judgment; the standard psychophysical measure of binocular depth sensitivity.

Stereopsis.

The perception of depth computed by the brain from binocular disparity; the percept to which vision disparity is the input.

Uncrossed disparity.

The disparity sign produced by an object farther than the fixation plane; signals that the object is beyond the point of convergence.

Key Researchers

Bruce G. Cumming

. Laboratory of Sensorimotor Research, National Eye Institute, NIH; demonstrated that primary visual cortex neurons respond to binocular disparity even when it produces no depth percept, separating the early encoding of the cue from the later computation of stereoscopic depth. ORCID

Gregory C. DeAngelis

. University of Rochester; co-developed the disparity-energy model of binocular neurons and showed by microstimulating area MT that extrastriate disparity signals causally drive stereoscopic depth judgments. ORCID

Béla Julesz

(1928-2003). Bell Labs; Rutgers University. Invented the random-dot stereogram, proving that binocular disparity alone, with no monocular form or familiarity cue, is sufficient for depth, and framed stereopsis as a correspondence problem the brain solves before recognizing any object. Wikipedia

Andrew J. Parker

. University of Oxford; synthesized how binocular disparity signals in the cerebral cortex are transformed into perceived depth, and with Cumming showed that V1 encodes absolute disparity without signalling the depth percept itself. ORCID

Gian F. Poggio

(1927-2007). Johns Hopkins University School of Medicine. Recorded the first disparity-selective neurons in the primary visual cortex of the awake, behaving monkey, including near, far, and tuned-zero cells, establishing the cortical substrate for stereopsis in the primate.

Jenny C. A. Read

. Newcastle University; advanced the computational and comparative understanding of stereopsis, modelling how the visual system solves correspondence and extending stereo research across the animal kingdom, including the discovery of stereoscopic vision in the praying mantis. ORCID

Andrew E. Welchman

. Wellcome Trust; formerly University of Cambridge. Used human neuroimaging and psychophysics to show how the brain reads disparity out into a three-dimensional percept, and co-identified “what not” detectors that reject false binocular matches. ORCID

Frequently Asked Questions

What is vision disparity?

Vision disparity, also called binocular or retinal disparity, is the small difference between the images the two eyes receive of the same scene. Because the eyes are about 6.5 centimetres apart, each views the world from a slightly different position, so any object nearer or farther than the fixation point projects to slightly different places on the two retinas. That angular offset is the disparity, and it is the geometric cue the brain uses to compute depth.

How is vision disparity different from depth perception?

Disparity is a cue; depth perception is the result. Disparity is a measurable difference in the two retinal images, whereas depth perception is the brain's final estimate of three-dimensional layout, built from disparity together with other cues such as motion, occlusion, and perspective. The Medical Subject Headings vocabulary files vision disparity beneath depth perception for exactly this reason.

What is stereopsis?

Stereopsis is the perception of depth that the brain computes from binocular disparity. It is what is lost when one eye is closed: the vivid, immediate sense of solidity and of the relative distances of nearby objects. Disparity is the input to stereopsis, and stereopsis is the percept it produces.

What is the correspondence problem?

The correspondence problem is the challenge of deciding which point in the left eye's image is the same world point as a given point in the right eye's image. Only matched points yield a true disparity, and in a scene full of similar features most possible pairings are wrong. The brain must solve this matching problem before it can measure disparity at all.

What is a random-dot stereogram?

A random-dot stereogram is a pair of fields of random dots, each meaningless on its own, in which a region of one image is shifted horizontally to create a disparity. When the two images are fused, the shifted region appears as a surface floating in depth. Béla Julesz invented it to prove that disparity alone, with no shape or familiarity cue, is enough to see depth.

Why is stereo vision so much better up close?

The disparity produced by a fixed depth difference falls off with the square of the viewing distance. A small depth step on a desk produces a large, easily seen disparity, but the same step across a room produces almost none. This inverse-square geometry makes binocular disparity a precise depth sense for the space within arm's reach and a weak one at a distance.

Do neurons that respond to disparity see depth?

Not by themselves. Neurons in primary visual cortex respond to disparity even in anticorrelated stereograms that produce no sense of depth, showing that encoding the cue is separate from perceiving depth. The depth percept emerges at later stages: stimulating disparity-tuned cells in area MT shifts an animal's depth judgments, tying that area to the percept.

Do animals other than humans have stereo vision?

Yes. Stereopsis has been demonstrated in a wide range of animals, including monkeys, cats, birds of prey, and even the praying mantis. Different species use different eyes and neural circuits but exploit the same geometric disparity cue, which is why comparative work helps separate the general principles of stereopsis from the particular way the mammalian brain implements it.

References