Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Peripheral Vision and Crowding

Abstract. Peripheral vision samples a broad visual field while restricting the recognition of individual objects in clutter. Visual crowding denotes the loss or distortion of target identity when surrounding elements interfere despite adequate isolated-target visibility. This fundamental research monograph traces the historical development of peripheral letter-recognition experiments, examines eccentricity and critical spacing, and distinguishes crowding from acuity, masking, localization errors, and decisional limitations. A critical narrative synthesis evaluates psychophysical thresholds, spatial anisotropy, feature similarity, statistical pooling, positional averaging, texture statistics, peripheral metamers, attentional selection, perceptual grouping, recurrent segmentation, and retinotopic neural constraints. Original empirical research is integrated with recent findings on the association between visual area V4 and crowding distance and developmental differences in foveal crowding. Separate analyses examine reading speed, visual search, eye movements, low-vision applications, observer differences, and controlled approaches to ecological validation. The evidence supports systematic interactions between spatial encoding, organization, and object-specific readout while leaving causal localization and cross-task generalization unresolved. The study identifies measurements and experimental contrasts capable of distinguishing theoretical accounts without treating a narrative literature synthesis as a formal systematic review.

Keywords: peripheral vision; visual crowding; Bouma's law; critical spacing; eccentricity; psychophysics; letter recognition; object recognition; positional averaging; peripheral metamers; texture statistics; perceptual grouping; retinotopic cortex; visual search; reading; eye movements; Artificial Era

Peripheral Vision and Crowding as a Scientific Object

Peripheral vision is the set of visual operations carried out at retinal locations away from the direction of gaze. It samples an extended field for orientation, motion, layout, objects, and prospective actions while resolving fine identity with less spatial precision than central vision. Visual crowding is a measurable impairment in identifying a target when nearby elements occupy critical neighboring positions, even when the same target can be recognized in isolation. The operative distinction is between visibility of constituent signals and availability of an individual object identity. A visual system may detect a peripheral target yet fail to identify its letter, orientation, or configuration because adjacent structure alters the representation used for recognition Whitney & Levi (2011) Rosenholtz (2016).

The research object combines stimulus geometry, neural sampling, feature interactions, recognition decisions, and naturalistic behavior. Eccentricity denotes angular distance from the fixation point; critical spacing is a target-to-flanker separation at which performance reaches a prespecified criterion. Acuity concerns the smallest resolvable detail, while crowding concerns the spatial isolation necessary for identifying a target among neighboring items. They are experimentally separable because enlarging a target can improve its single-item acuity without removing interference from flankers. The familiar experience of seeing text as present but being unable to name an individual peripheral character captures this distinction without serving as its operational proof Pelli & Tillman (2008) Levi (2008).

This monograph is a critical narrative synthesis of psychophysics, behavioral experiments, mathematical models, neurophysiological measurements, and applied studies. It asks how peripheral representations encode object information, which manipulations identify crowding independently of masking, how models predict systematic identification errors, and what can legitimately be inferred from visual cortical measurements. Evidence from controlled letter arrays is connected to natural scenes and reading, with direct recognition that external validity remains an empirical question. The focal scope is peripheral recognition and interference; comprehensive histories of object perception, visual attention, and central acuity belong to neighboring research objects Strasburger et al. (2011) Schwetlick & Herzog (2025).

The Foveated Geometry of Human Visual Sampling

Human visual sampling is highly nonuniform. An observer directs the fovea toward selected details while the surrounding retina supplies a wider, lower-resolution field. Optics, retinal cell distribution, cortical magnification, receptive-field organization, and behavioral selection all contribute to spatial variation. Yet the decline of isolated-target acuity does not determine the decline of recognition within clutter. Crowding exposes a functional bottleneck at larger spatial scales: information about multiple nearby items can be present in a pooled or confused representation even when object-specific readout is poor. The review by Strasburger, Rentschler, and Jüttner distinguishes peripheral pattern recognition from a simplistic account of visual blur Strasburger et al. (2011).

A peripheral display must be specified in degrees of visual angle rather than by pixel coordinates alone. Viewing distance, target size, center-to-center separation, fixation location, and display calibration determine whether an experiment tests the proposed retinal geometry. A target at a fixed screen position changes its eccentricity if gaze drifts. Eye tracking is therefore especially important when crowding is measured near the fovea, where small errors in fixation become consequential. Vertical and horizontal meridians should be analyzed rather than averaged automatically, since spatial interactions and recognition performance can be anisotropic Toet & Levi (1992) Greenwood et al. (2017).

The phenomenology of peripheral vision is likewise spatially structured. An observer can apprehend the overall character of a street or crowded page while failing to individuate a small peripheral object. This pattern is consistent with retention of aggregate information amid local losses, but successful scene categorization does not prove one particular compression algorithm. The experimenter must compare models against discriminative observations such as error distributions, target-flanker similarity, flanker count, and grouping structure. These measurements help separate the descriptive statement that information is lost from a causal theory explaining where and why that loss occurs Balas et al. (2009) Rosenholtz et al. (2012).

From Parafoveal Letter Recognition to Contemporary Crowding Science

Systematic investigations of recognition outside the fovea preceded the contemporary vocabulary of crowding, but Herman Bouma's 1970 report on parafoveal letter recognition became an enduring point of reference. Bouma studied the effects of nearby letters on recognizing an eccentric target and connected target-flanker interference to visual reading. Subsequent work formalized a relation between required spacing and eccentricity, usually referred to as Bouma's law. The historical value of the paper is its specific experimental contribution to identifying spatial interactions, not a timeless assertion that every crowding task follows one universal proportionality Bouma (1970) Strasburger (2020).

Experiments in the 1990s replaced simple verbal generalization with mapping of interaction zones. Toet and Levi measured interference for oriented T-like elements at several eccentricities and meridians, showing that crowding occupies a two-dimensional and direction-dependent region around a target. The localization of interference matters: a single radius cannot describe all possible flanker arrangements in every part of the visual field. Work on orientation adaptation also suggested that early feature processing survives conditions in which observers cannot explicitly identify the crowded target, motivating explanations based on later selection or integration stages Toet & Levi (1992) He et al. (1996).

During the 2000s, researchers made crowding a more explicitly mechanistic field by contrasting it with ordinary pattern masking and by analyzing errors rather than thresholds alone. Pelli, Palomares, and Majaj examined conditions in which target detection and identification diverged, arguing that crowding concerns inappropriate combination of features rather than mere disappearance of visual energy. Theoretical reviews subsequently extended the problem to faces, reading, visual search, and scene perception. Current accounts include pooling, substitution, population coding, grouping, recurrent segmentation, and image-statistical compression, with continued debate over causal localization Pelli et al. (2004) Pelli & Tillman (2008) Herzog et al. (2015).

Separating Crowding, Acuity, Masking, and Source Confusion

An operational crowding experiment measures identification of a defined target with and without flanking objects under controlled visibility. The distinction from acuity depends on showing adequate isolated-target performance at the same location, size, contrast, and exposure. The distinction from masking is harder because flankers can alter both low-level signal strength and higher-level segmentation. In a robust design, researchers vary contrast, spacing, stimulus class, or temporal presentation and compare the resulting signatures against predictions of alternative mechanisms. Reduced identification accuracy is an observation requiring explanation, not a sufficient criterion for choosing one mechanism Pelli et al. (2004) Chung et al. (2001).

Source confusion occurs when observers report the identity of a neighboring flanker rather than the designated target; feature substitution similarly describes errors with the character of a competing item. Such responses may reveal how the brain assigns features to spatial positions, although choosing a flanker is also compatible with uncertainty in the response stage. Averaging errors, by contrast, can produce judgments intermediate between target and flanker values. A single percentage-correct score discards this structure. Continuous-report distributions, confusion matrices, and carefully counterbalanced flanker identities offer stronger tests of pooling and substitution explanations Greenwood et al. (2009) Harrison & Bex (2015).

Terminology itself deserves precision. Some studies call a stimulus crowded whenever flankers impair recognition; others reserve the term for interference whose spatial scaling differs from conventional masking. The literature also distinguishes letter-string crowding, foveal crowding, object crowding, and more general contextual interference. Investigators need to state their threshold criterion and whether the quoted spacing is center-to-center or edge-to-edge. Strasburger's historical reanalysis shows why seemingly standard statements about Bouma's law, periphery, and foveal limits can conceal incompatible operational conventions Strasburger (2020) Levi (2008).

Critical Spacing, Eccentricity, and the Bouma Relation

Bouma's law describes an approximate growth of the critical center-to-center spacing with target eccentricity. A common shorthand gives critical spacing of roughly one half of eccentricity for canonical peripheral letter tasks, but the numerical factor varies with stimuli, criterion, meridian, and task. It is best viewed as a constrained empirical regularity rather than a universal physiological constant. The equation s approximately equals b times e is useful for design when s is spacing, e is eccentricity, and b is estimated for the actual observer and paradigm. Intercepts near central vision and nonlinearity at fine eccentricities require separate consideration Pelli & Tillman (2008) Strasburger (2020).

Spacing measurements often reveal a stronger limitation than retinal resolution. If a letter is enlarged above its acuity threshold without proportionally separating its neighbors, crowded recognition can remain deficient. Conversely, separating neighboring letters can improve identification while the target's physical size remains unchanged. This logic underlies the uncrowded-window account, which treats the number of separately recognizable objects as constrained by available visual field geometry. The account explains broad regularities but must be qualified when grouping, target-flanker similarity, or retinal location alters interference beyond the simplest proportional rule Pelli & Tillman (2008) Manassi et al. (2012).

The method used to estimate a critical distance shapes interpretation. Psychometric functions estimated from target identity reports at multiple separations may yield different thresholds according to lower asymptote, lapse rate, response alternatives, and fitting method. When flankers differ in size from the target, center-to-center and edge-to-edge metrics diverge; when observers move their eyes, nominal and retinal separation diverge. A generalizable claim therefore requires reporting both behavioral criterion and geometry. The two-dimensional maps of Toet and Levi and subsequent field-topology studies illustrate why average critical spacing can obscure systematic individual and meridional structure Toet & Levi (1992) Greenwood et al. (2017).

Anisotropy, Radial Organization, and Inner–Outer Asymmetry

Crowding zones can be elongated radially relative to fixation, and a flanker positioned farther from fixation may exert different interference from an equally distant flanker nearer the fovea. These effects imply that a visually symmetric array need not be functionally symmetric. Meridional differences add another dimension: thresholds may differ along upper, lower, nasal, and temporal directions. The traditional sketch of a circular zone around every target is therefore an inadequate quantitative specification. Toet and Levi directly mapped interaction shapes; later work linked individual spatial variation across recognition and localization tasks Toet & Levi (1992) Greenwood et al. (2017).

Inner–outer asymmetry is particularly sensitive to how an experiment defines distance and arranges neighboring characters. In a line of letters, outer and inner positions differ in eccentricity, projected cortical magnification, and possible grouping relationships. A design that manipulates only flanker distance without separately balancing retinal location may conflate these contributions. Mechanistic interpretation requires checking the model against multiple orientations and controlling target identity, contrast, exposure, and gaze. The observed geometry constrains theories that posit fixed isotropic receptive fields or a single all-purpose attentional window Strasburger (2020) Nandy & Tjan (2012).

Anisotropy also affects application. A text layout optimized using a single global value for minimum interletter spacing may serve poorly across gaze directions or observers. Human-factors models should specify likely fixation points, target locations, display size, movement demands, and whether peripheral recognition is actually required for the task. The experimental aim is to estimate behavior in that context rather than to transfer a mean parameter obtained from unrelated letter arrays. Spatial variability should be treated as meaningful data within both psychophysics and ergonomics Rosenholtz (2016) Greenwood et al. (2017).

Experimental Designs and Psychophysical Identification Tasks

Peripheral crowding is most commonly measured by presenting a target at a controlled eccentricity and manipulating the locations or attributes of nearby flankers. Letters, Landolt rings, oriented gratings, faces, contours, and synthetic textures each sample different representational requirements. Forced-choice identification reduces ambiguity about scoring but introduces guessing rates and category restrictions. Continuous reports of orientation, position, or color expose response distributions that a simple correct-versus-incorrect score would suppress. The central inferential question is whether a parameter describes sensory encoding, target individuation, decision uncertainty, or their joint outcome. Comparisons must preserve isolated-target discriminability Pelli et al. (2004) Harrison & Bex (2015).

Trial duration, eye position, and stimulus timing are central design variables. Brief presentations and fixation monitoring reduce the possibility that observers foveate the target before responding. Eye tracking enables removal or modeling of trials with anticipatory saccades, but an eye tracker requires calibration and explicit thresholds for permissible fixation drift. Randomized target and flanker identities prevent learning a predictable neighborhood. Spatial uncertainty, response mapping, display contrast, and target visibility should be treated as design factors rather than incidental implementation details Strasburger et al. (2011) Greenwood et al. (2017).

The strongest investigations combine complementary measures. Critical-spacing functions establish where recognition changes, unflanked psychophysics establishes what can be detected, response-distribution measures reveal the form of errors, and task manipulations test hypotheses about selection and grouping. Reliability should be assessed across sessions because estimated thresholds may depend on attention, fatigue, and adaptation. Researchers should publish stimulus geometry and enough code or trial data for independent reproduction. A convincing account predicts condition-specific patterns rather than explaining every observed decrement as crowding after the fact Schwetlick & Herzog (2025) Strasburger (2020).

Psychometric Functions and Parameter Identification

A psychometric function relates a controlled stimulus variable, such as target–flanker separation, to the probability of correct recognition. Its lower asymptote depends on the number of response alternatives and possible response biases; its upper asymptote can be reduced by inattentive lapses or genuine residual error. A threshold may be defined at a criterion performance level or relative to estimated asymptotes, and different definitions produce nonidentical values. Fitted slope carries information about the transition from strong to weak interference. Reporting only a single threshold masks model uncertainty and prevents comparison across experiments using different criteria Chung et al. (2001) Pelli et al. (2004).

Adaptive procedures allocate trials near the estimated threshold and can reduce testing burden. They may also sample a narrow range of separations, making lapse-rate estimation unstable or reducing opportunities to detect multiple transitions. Fixed-level sampling is less efficient but offers stronger direct visualization of the response function and residuals. Hierarchical fits can estimate observer-level effects without pretending that every trial is independent evidence for a population conclusion. Formal comparison should include uncertainty in eccentricity, scale, target identity, and session effects when those quantities vary Greenwood et al. (2017) Kurzawski et al. (2025).

Psychophysical measurement is especially vulnerable to unit confusion. Angular spacing, pixel spacing, relative target diameter, and cortical millimeters refer to different quantities. The 2025 V4 study uses a derived count of uncrowded letters and measured surface areas of retinotopic maps, so its cortical spacing inference must be interpreted within its adopted geometry and scaling assumptions. A threshold from one physical representation cannot be imported into another without transformation. A useful experimental report supplies raw viewing parameters, definitions, confidence intervals, psychometric form, exclusion rules, and diagnostic plots Kurzawski et al. (2025) Strasburger (2020).

Letter Recognition and the Boundary of Identifiable Form

Letter identification provided some of the earliest persuasive demonstrations of crowding because symbols can be physically simple and yet require accurate conjunctions of strokes, orientation, and position. A recognizable single letter becomes difficult to identify when closely surrounded by other characters. Target location should be explicitly cued; otherwise errors can reflect selecting a different letter from the array. The experimenter can vary flanker similarity, character spacing, font, case, contrast polarity, and exposure. Bouma's original parafoveal observation and later demonstrations show that success at detecting a character is distinct from success at assigning its identity Bouma (1970) Pelli et al. (2004).

The use of letter strings supports precise response scoring but creates familiar-language and linguistic-context effects. Letter frequency, word shape, positional expectation, and reading expertise may alter report or recognition. Random trigrams reduce higher-level lexical support, whereas meaningful words invite contextual prediction. Some apparent improvements may arise from choosing a likely letter rather than recovering its visual features. Studies must specify whether the goal is character-level discrimination, word recognition, or continuous reading. These are related outcomes with different perceptual and linguistic components Chung et al. (1998) Chung et al. (2001).

Letter tasks also reveal why crowding cannot be reduced to a general inability to perceive peripheral detail. A subject may detect a small isolated symbol at an eccentric location while failing to identify a larger target in an array. Increasing character size alone is not equivalent to increasing center-to-center separation. The uncrowded-window model highlights spacing as an independent visual limit; alternative accounts emphasize which letters group together or how source positions become corrupted. Both can be examined by analyzing exactly which competing characters are reported Pelli & Tillman (2008) Greenwood et al. (2009) Herzog et al. (2015).

Feature Similarity, Contrast, and Cross-Dimensional Interference

Target–flanker similarity can alter the severity of crowding. If adjacent items differ in color, contrast polarity, orientation, shape, or depth, they may segregate more readily and produce less interference. The size of this release depends on visual task and physical visibility, so a similarity effect alone cannot identify a neural stage. A mismatch can change low-level filtering, perceptual grouping, and decision strategy at once. Psychophysical studies therefore compare combinations of similarity dimensions and preserve target discriminability across conditions whenever possible Chung et al. (2001) Herzog et al. (2015).

Chung, Levi, and Legge manipulated spatial-frequency content and contrast in letter crowding, placing an early constraint on treating it as ordinary lateral masking. Other experiments use orientation averaging or distinctive surrounds to show that the interference may operate on feature estimates rather than simply reducing sensitivity. Dakin and colleagues separated effects of attention and crowding in orientation averaging, establishing that related task deficits need not reflect a single attentional cause. The explanatory value of these demonstrations lies in their pattern of dissociations, not just in the existence of a significant flanker effect Chung et al. (2001) Dakin et al. (2009).

Feature-specific findings raise a central question: is crowding one mechanism expressed across visual attributes, or a family of interactions sharing a spatial signature? Orientation, color, motion, letter identity, and face recognition involve partially distinct representational requirements. A common critical-spacing trend does not establish identical computational circuitry. Conversely, differences in individual tasks do not show complete mechanistic independence. The most informative tests assess multiple feature dimensions within the same observers and explicitly compare unified versus domain-specific model predictions Whitney & Levi (2011) Schwetlick & Herzog (2025).

Location Errors, Averaging, and Competing Sources of Responses

Errors in a crowded display can reveal how multiple inputs are combined. In a continuous orientation task, a report may shift toward flankers, cluster around one flanker value, or broaden without a systematic directional bias. These outcomes are distinguishable with response distributions but may look identical under binary scoring. Greenwood, Bex, and Dakin demonstrated how positional averaging can explain effects with letter-like stimuli. This suggests that confusion over position can be as important as impairment in coding feature identity Greenwood et al. (2009).

The phrase feature substitution often refers to reporting a flanker property as if it belonged to the target. It is a description of a response pattern that still requires a mechanism. A noisy assignment of features to objects, a probabilistic competition between candidate positions, or a decoding rule that weights neighbors could generate similar substitutions. Computational models must predict the full distribution of errors across flanker values and spatial configurations. The population-code approach of van den Berg and colleagues and the 2015 unifying orientation model provide contrasting formal means of doing so van den Berg et al. (2010) Harrison & Bex (2015).

Interpreting forced-choice data requires care because a target can have multiple feature dimensions and the experimenter's response set may privilege one. A target location cue may reduce source ambiguity without changing the pooled visual encoding. Response biases may change with prior probabilities of target identity, and prolonged viewing can recruit compensatory search or eye movements. Strong causal claims therefore require controls for response selection, perceptual visibility, and the amount of visual information physically present. Combining identity and localization measures is preferable to inferring object representations from identification accuracy alone Greenwood et al. (2017) Pelli et al. (2004).

Perceptual Grouping, Gestalt Organization, and Release From Crowding

Grouping is an influential challenge to models based only on local distance. Surrounding elements sometimes form a perceptually coherent group that segregates from the target, improving target identification despite increased element count. Conversely, a target that becomes grouped with flankers may be more difficult to isolate. Manassi, Sayim, and Herzog studied such configurations in which bigger arrays can be better, demonstrating that adding elements need not produce monotonically increasing interference. These results require models to account for global arrangement as well as local spacing Manassi et al. (2012).

Grouping mechanisms can be expressed through similarity, continuity, symmetry, regularity, common region, and completion. Some experiments demonstrate release from crowding when a global contour or repeated pattern changes the target's relation to its surroundings. Herzog and colleagues argue that visual appearance and object organization matter to the phenomenon itself, suggesting a dynamic segmentation process. Yet a grouping explanation remains incomplete unless a model defines which spatial operations produce the inferred object boundaries and can predict individual trial performance Herzog et al. (2015) Francis et al. (2017).

An especially useful experimental strategy constructs matched displays with similar local target–flanker configurations but different global layouts. If identification changes, a purely local interference rule is insufficient. Equally, a recurrent grouping model must explain classic spatial-scaling and feature-error phenomena rather than treating them as exceptions. Models need to show how a representation carries information at different organizational scales and how the target is separated from its neighbors. The growing literature suggests multiple interacting determinants whose relative weight depends on task and stimulus configuration Schwetlick & Herzog (2025) Manassi et al. (2012).

Feature Pooling and Compulsory Integration Models

Pooling theories explain crowding as integration of target and neighboring features within regions whose size increases outside the fovea. A decoder accessing pooled orientation or stroke information can recover that features are present while losing the binding required to assign them to a particular object. Compulsory integration predicts systematic biases rather than entirely random errors, including reports intermediate between the target and flanker values. The framework gained support from studies of orientation averaging and other continuous feature judgments. However, the word pooling denotes a family of mathematical operations; a viable model must specify receptive-field shape, weighting, normalization, and readout Levi (2008) van den Berg et al. (2010).

A simplistic average of all nearby elements fails to explain every grouping effect. If additional flankers cause a coherent surround to segregate from the target, recognition may improve even though the number of available features has increased. Pooling models may incorporate grouping-dependent weights, but doing so introduces parameters that require independent measurement. The important comparison is whether those parameters predict novel layouts or simply accommodate an observed result. Directly contrasting spatial-only pooling with object-contingent weighting can expose genuine theoretical differences Manassi et al. (2012) Herzog et al. (2015).

Pooling is sometimes discussed as though it entails a permanent loss of all local detail. Behavioral evidence supports task-dependent failures of readout, while orientation adaptation and other implicit responses may persist under crowding. A representation can retain information that is inaccessible to a specific forced-choice report; alternatively, it can preserve a coarse statistic while discarding fine spatial labels. Experiments must therefore identify the measured level of processing before equating performance with information loss. A scientifically productive model makes both recognition-error and residual-information predictions He et al. (1996) Freeman & Simoncelli (2011).

Positional Averaging and Spatial Uncertainty

Spatial uncertainty concerns the reliability with which visual features are assigned positions. It is a plausible contributor to crowding because small errors in feature localization can transform the perceived geometry of a letter even when individual strokes remain visible. Greenwood, Bex, and Dakin's letter-like stimuli supplied evidence that averaging relative positions can account for some crowding errors. This approach shifts the emphasis from deleting feature representations to corrupting the relations between otherwise encoded elements Greenwood et al. (2009).

Spatial uncertainty can be tested by varying positional offsets, asking observers to localize separate components, and comparing response biases with candidate integration rules. A Bayesian or other probabilistic estimator might place more weight on neighbors when the target's location is poorly specified. Yet averaging location does not necessarily explain all category recognition errors, especially where global grouping or semantic object identity matters. The model should be evaluated at fixed eccentricity across multiple arrangements and against trials that preserve the same local feature statistics but alter spatial organization Greenwood et al. (2017) Herzog et al. (2015).

Visual-field topology provides additional constraints. Greenwood and colleagues compared crowding zones, saccadic precision, and spatial-localization measures and found both shared idiosyncratic variation and dissociable task effects. Their results support correlated spatial organization without reducing all behaviors to one common representation. This distinction is important in causal reasoning: a correlation between location uncertainty and crowding may reflect inherited field structure rather than a direct single-stage cause. Tests of mediation require manipulations designed to alter one component while monitoring the others Greenwood et al. (2017).

Summary Statistics and Peripheral Texture Representations

Summary-statistic theories treat peripheral vision as a compression system that retains aggregate features of image regions while sacrificing some precise spatial arrangement. Balas, Nakano, and Rosenholtz proposed joint statistics of local filter responses across orientation, position, scale, and phase to account for several crowding phenomena. Such representations can support scene gist, texture characterization, and guidance of fixation, even when an isolated target cannot be reconstructed. The approach links object crowding to a broader computational account of the information available outside central vision Balas et al. (2009) Rosenholtz (2016).

The theory can be tested by creating images matched on modeled summary statistics yet differing in fine arrangement. If observers frequently confuse such images in the periphery, the model captures a constraint on accessible visual information. The strength of this strategy is that model variables imply specific stimulus equivalence classes. Its limitation is that perceptual indistinguishability depends on display geometry, observer, task instructions, and a chosen representation; a statistical match is not automatically a proof of neural implementation. Rosenholtz and colleagues showed how reduced peripheral information can mimic some effects attributed to attention during visual search Rosenholtz et al. (2012).

Summary-statistic explanations must also engage grouping data. If global arrangement determines whether target and flankers belong to a common object, then fixed local windows may require adaptive segmentation or context-sensitive weighting. Adding mechanisms can improve descriptive range but makes model comparison more demanding. A meaningful account should predict when observers perceive texture accurately, when they recognize a particular object, and when the same array supports one task but defeats another. The empirical target is a structured loss of identity and position, not an assumption that the entire peripheral image is uniformly degraded Schwetlick & Herzog (2025) Balas et al. (2009).

Peripheral Metamers and the Hierarchy of Visual Representation

Metamers are different physical stimuli that produce indistinguishable percepts under specified viewing conditions. Freeman and Simoncelli constructed model-based visual metamers by matching nonlinear combinations of early filter responses pooled over spatial regions, and used psychophysics to constrain the spatial scale of representation. Their 2011 experiments implicated a level comparable to visual area V2 for the tested metamers, showing that model-generated equivalences can probe more than simple acuity limits Freeman & Simoncelli (2011).

The metameric method is powerful because it converts a representational theory into a falsifiable comparison. A model that predicts equivalence must produce stimuli the observer fails to distinguish at fixed eccentricity and task. If participants reliably discriminate candidate metamers, some information missing from the model remains accessible. Spatial windows, filter statistics, and viewing time must be controlled; model success with textures does not automatically demonstrate success with letters or faces. The appropriate inference is a correspondence between tested behavioral equivalence and a hypothesized class of neural computations Freeman & Simoncelli (2011) Rosenholtz (2016).

Modern image-generative models can produce convincing peripheral approximations, but perceptual similarity judged at fixation is not an adequate validation. The visual system's effective encoding changes with eccentricity, fixation, and viewing task. A generated surrogate that succeeds in a scene-gist judgment may fail in target localization or letter discrimination. Likewise, an image-processing algorithm that reproduces a crowding curve does not establish that cortex uses its exact code. Model-based metamers should be compared with measured confusion matrices, controlled eye tracking, and alternative image statistics to constrain explanation Balas et al. (2009) Schwetlick & Herzog (2025).

Attention, Resolution, and the Accessibility of Peripheral Features

Attention may alter crowded recognition by prioritizing a location, enhancing feature precision, or selecting one representation for decision. He, Cavanagh, and Intriligator reported orientation-specific adaptation surviving conditions where crowded orientation was unavailable for explicit report. They interpreted this dissociation as evidence for limited attentional resolution beyond an early cortical coding stage. The result is historically influential, although adaptation survival and impaired report do not uniquely locate the bottleneck to attention; pooling, segmentation, and response access provide further hypotheses He et al. (1996) Whitney & Levi (2011).

Attentional allocation and crowding can be manipulated separately. A valid location cue may improve target performance, yet residual crowding can remain; a feature-report task may respond differently from detection. Dakin, Bex, Cass, and Watt reported dissociable effects of attention and crowding on orientation averaging. Consequently, all attention-dependent changes in performance should not be relabeled as changes in crowding strength. Researchers need factorial designs varying spacing, cue validity, exposure, task difficulty, and target detectability, along with explicit tests for interactions rather than relying on one main effect Dakin et al. (2009).

Rosenholtz, Huang, and Ehinger modeled visual-search behavior in terms of the information lost in peripheral encoding, challenging interpretations that attribute every search limitation to a mobile attentional spotlight. This position does not imply that selection and attention are irrelevant. It imposes a stronger requirement: demonstrate what attention explains after the constraints of fixed eccentricity-dependent representation have been modeled. The productive question concerns how peripheral encoding and selection jointly determine fixation planning, target identification, and awareness under different environments Rosenholtz et al. (2012) Rosenholtz (2016).

Segmentation, Recurrent Dynamics, and Contextual Integration

Object organization modifies the spatial field over which visual information is integrated. In recurrent accounts, feedforward feature extraction interacts with feedback or lateral connections that segment target and flanker groups over time. Francis, Manassi, and Herzog developed a neural-dynamics model explaining several grouping and crowding properties, extending theories beyond a fixed local aggregation window. This class of model treats segmentation as an operation whose behavior depends on arrangement, similarity, and the temporal evolution of activity Francis et al. (2017).

The distinction between local and global effects can be tested by changing elements outside a classical crowding zone while preserving nearby target features. If global completion releases a target from crowding, a strictly independent local window is challenged. However, a recurrent model must also predict familiar spacing laws and specific error distributions. Global grouping effects are sensitive to the arrangements chosen, and researchers should inspect whether local feature energy, attention, or response uncertainty changed inadvertently. Theory comparison depends on matched inputs and quantitative fits to held-out conditions Manassi et al. (2012) Herzog et al. (2015).

Cicchini, D'Errico, and Burr reported crowding patterns consistent with statistically efficient integration of target and contextual information. Their 2022 account highlights that combining signals can be adaptive under uncertainty even when it appears disadvantageous in a target-isolation task. This is a mechanistic hypothesis with testable predictions about weighting and uncertainty, not proof that every crowded error optimizes a single objective. Future experiments can compare integration weights across tasks, naturalistic priors, and changing reliability while establishing whether observers actually acquire the assumed statistics Cicchini et al. (2022) van den Berg et al. (2010).

Cortical Maps, Neural Coding, and Limits of Localization

Neural explanations seek mechanisms connecting retinal eccentricity to the spatial scale of recognition. Retinotopic maps represent visual locations across cortical tissue, and cortical magnification differs across the visual field. A parsimonious hypothesis is that objects need sufficient separation in a relevant cortical map for their features to remain independently readable. Such reasoning produces predictions about the relation between measured crowding thresholds and cortical surface area. It is more informative than assigning the entire effect to an unnamed higher visual center because it specifies which brain measurements should covary with behavior Nandy & Tjan (2012) Kurzawski et al. (2025).

Kurzawski and colleagues measured crowding psychophysically and retinotopic areas with fMRI in 49 observers. The peer-reviewed 2025 study found that their derived measure of uncrowded letter capacity scaled with area V4 and was not similarly related to areas V1–V3. The fitted conservation model implied a cortical spacing of approximately 1.4 millimeters in V4, with model assumptions and uncertainty reported by the investigators. Correlation and model fit constrain possible neural organization; they do not demonstrate that V4 is uniquely necessary for crowding or that one neuronal mechanism accounts for all stimuli Kurzawski et al. (2025).

An alternative account from Nandy and Tjan linked crowding to learning image statistics under saccade-related confounds. It illustrates a causal-developmental possibility that connects eye-movement history with peripheral neural structure. The scientific challenge is to distinguish developmental learning, current cortical geometry, and moment-to-moment perceptual integration with interventions capable of separating their predictions. Functional imaging alone has limited temporal and causal resolution; lesion evidence, neural perturbation, electrophysiology, and longitudinal measurement can supply complementary constraints. A localized correlate does not settle how the visual system binds target features under clutter Nandy & Tjan (2012) Greenwood et al. (2017).

Foveal Crowding and Development Across Childhood

Crowding is especially prominent outside the central visual field, but an absolute claim that it cannot occur near the fovea is empirically false. Near-central measurements require precise fixation, careful geometric definitions, and discrimination from overlap masking. In children, foveal crowding can be substantially greater than in mature observers, raising questions about the developmental refinement of object individuation and segmentation. The strength of conclusions depends on matching task comprehension and isolated-target acuity across ages, not merely comparing raw flanked performance Strasburger (2020) Greenwood et al. (2026).

Greenwood, Kyprianou, and Dekker studied foveal gap-orientation discrimination in a final sample of 119 participants, comprising typically developing children and adults, using isolated and flanked stimuli. Their study, published online in December 2025 in the 2026 volume of Scientific Reports, found that isolated acuity approached adult levels earlier than flanked recognition. Flanker-related threshold elevations remained conspicuous at younger ages and reached adult-like levels later in the tested groups. Their synthesis of 13 prior studies emphasized the importance of differing stimulus geometry and threshold conventions when interpreting developmental timing Greenwood et al. (2026).

Developmental results should not be transformed into a universal maturation age or used as a diagnostic test without clinical validation. The distribution of ages, practice, motor response, visual acuity, classroom experience, and instructions all affect comparisons. Importantly, similarity-dependent release from interference in younger observers provides a way to test continuity between childhood foveal and adult peripheral mechanisms. Longitudinal data and cross-task measurement invariance would distinguish maturation of perceptual encoding from improved response control or attention, a distinction that cross-sectional age comparisons cannot fully resolve Greenwood et al. (2026) Schwetlick & Herzog (2025).

Crowding, Reading, and the Visual Span

Reading requires rapid identification of multiple adjacent letter forms while gaze advances through saccades. Crowding is therefore a natural candidate mechanism limiting the useful visual span, especially when reading occurs outside the fovea. However, reading speed also depends on lexical recognition, eye-movement planning, attention, and language processing. A change in single-letter crowding does not establish the same proportional change in continuous reading. The research question is which component of reading is restricted by target-flanker interference and whether an intervention transfers from recognition tasks to text comprehension or fluency Pelli & Tillman (2008) Chung et al. (1998).

Chung, Mansfield, and Legge manipulated print size while measuring reading speed in normal peripheral vision. Their 1998 study helped establish that simply enlarging text does not necessarily recover foveal reading performance at eccentric positions. The design distinguished the size necessary for recognition from asymptotic speed under different retinal conditions. A reader using an eccentric retinal location may require larger print yet remain limited by other factors. This makes task-level performance a crucial outcome measure alongside character-specific thresholds Chung et al. (1998).

Training findings show the need for transfer tests. In 2007, Chung demonstrated that practice identifying crowded letters reduced crowding measures while improvements in peripheral reading speed were comparatively small in the tested experiment. Such a pattern is compatible with genuine perceptual learning in the trained task and with additional constraints on reading. The measured outcomes should be reported independently, including training dose and retention. Claims that reducing crowding necessarily produces a clinically important increase in reading speed extend beyond the evidence of that demonstration Chung (2007) Chung et al. (1998).

Visual Search, Scenes, and Ecological Recognition

Visual search begins from a large eccentric field in which candidate objects may be only partially identifiable. Peripheral information can nevertheless guide eye movements toward likely targets, meaning that task performance reflects both encoding and sampling strategy. Summary-statistic theories posit that a representation preserving local texture or feature distributions can support coarse guidance before individual object recognition. This account encourages explicit consideration of information sufficient for the current decision. A driver looking toward an intersection and a reader inspecting adjacent letters impose different requirements on the same broad visual capacity Rosenholtz et al. (2012) Rosenholtz (2016).

Ringer, Coy, Larson, and Loschky investigated crowding using complex rendered scenes and object categorization rather than only isolated letter arrays. Their 2021 experiments addressed whether the interference measured in simple displays transfers to cluttered real-world-like images, with carefully manipulated contextual objects. The approach increases ecological relevance while creating difficulties in equating target contrast, semantic context, and visual complexity. Experimental recognition in a synthetic scene remains a controlled approximation to dynamic real-world perception; its assumptions should be made explicit Ringer et al. (2021).

Attributing search difficulties solely to deficient top-down attention risks neglecting information that was never available at the required resolution in peripheral encoding. Conversely, explaining all search behavior through summary statistics neglects strategy, prior knowledge, and task incentives. A principled design predicts which image regions should be discriminable from a given fixation and then measures how gaze and response behavior change across conditions. Natural images, controlled noise, peripheral metamers, and recorded scan paths together permit stronger inference than any one laboratory task Rosenholtz et al. (2012) Freeman & Simoncelli (2011).

Saccades, Fixation, and Active Visual Sampling

The foveated visual system is active: observers repeatedly reposition gaze to obtain high-fidelity samples of selected parts of a scene. Peripheral recognition determines which locations appear informative enough to inspect, and saccades change the retinal location of the next target. Crowding can therefore affect action indirectly through information available for selecting a fixation rather than through a deficit in movement execution. Eye-movement experiments should distinguish fixation selection, saccade endpoint precision, and post-saccadic identification as separate dependent variables Greenwood et al. (2017) Rosenholtz (2016).

Greenwood, Szinte, Sayim, and Cavanagh found common spatial-field variations in crowding, saccadic precision, and localization, together with dissociations indicating distinct representations. A strong spatial correlation between tasks can reveal shared field topology while failing to show that crowding directly causes saccade imprecision. Interventions changing target visibility or manipulating flanker structure within observers can help test causal pathways. The principal methodological risk is reifying a correlated spatial map into a single psychological faculty without measuring distinct intermediate stages Greenwood et al. (2017).

Nandy and Tjan offered a provocative account in which statistics learned from visual experience could reflect artifacts of saccadic sampling, thereby generating crowding-like limitations. It motivates developmental studies and computational simulations that manipulate experienced image distributions or sampling policies. Such accounts need to explain classic crowding properties across tasks and observers, including grouping effects and anisotropies. Eye movements are therefore both a potential explanatory process over development and an immediate source of contamination in fixation-controlled psychophysics Nandy & Tjan (2012) Toet & Levi (1992).

Clinical, Low-Vision, and Accessibility Implications

Visual crowding matters for low-vision performance because a readable isolated symbol may become unrecognizable when embedded in ordinary text or crowded layouts. Differences between acuity-chart measures and real reading tasks can therefore be functionally important. Yet visual conditions associated with reduced acuity, unstable fixation, or abnormal eye movements introduce additional mechanisms, and controlled findings in normally sighted adults cannot be imported into clinical populations without direct evidence. Evaluation should identify the visual field region used, acuity, fixation stability, print geometry, cognitive demand, and task-level outcomes Chung et al. (1998) Strasburger (2020).

Perceptual training demonstrates both promise and specificity. A study in people with longstanding central vision loss found improvement in reading speed after practice, but such observations require careful attention to sample size, absence of a matched control, task practice, and durability of benefits. The earlier crowded-letter training work found stronger transfer to laboratory crowding indices than to peripheral reading speed. Together these studies support investigating training rather than prescribing a single crowding-targeted therapy as established treatment. Accessibility applications need outcomes meaningful to readers, not only psychophysical indices Chung (2007) Levi (2008).

Design recommendations can be generated as hypotheses from established mechanisms: provide adequate separation between critical symbols, allow gaze shifts, avoid densely packed peripheral alerts, and evaluate complex interfaces under realistic tasks. They require validation in the target population and context. The research does not justify numerical universal spacing rules, disease diagnosis from an unvalidated crowding chart, or promises of recovery through one intervention. Cross-disciplinary work with optometry, rehabilitation science, and human factors should preserve these evidence boundaries and test performance, comfort, and error severity directly Rosenholtz (2016) Ringer et al. (2021).

Evidence Hierarchy, Replication, and Model Comparison

The evidence base contains classic controlled experiments, modern psychophysics, computational modeling, neuroimaging, naturalistic scenes, developmental studies, and broad narrative reviews. These methods answer different questions. Experimental flanker manipulations can establish causal effects of local context under a specific task; correlations with cortical maps locate explanatory constraints; simulations demonstrate that a mechanism can reproduce observations; reviews integrate heterogeneous findings. None of these designs independently establishes a complete theory of crowding. Evidence evaluation should identify which observations are direct measurements, which depend on parameter estimation, and which arise from interpretive assumptions Whitney & Levi (2011) Schwetlick & Herzog (2025).

Replication requires more than obtaining a significant reduction in accuracy with flankers. Research should reproduce the geometry of the critical-spacing function, the shape of error distributions, anisotropies, effects of grouping, and individual variability, with independent participants and suitable stimuli. Competing accounts should be fitted to shared datasets and tested on held-out conditions. A model that fits classic letter trigrams may fail on contour completion or real-world objects. Good evidence synthesis preserves these failures because they identify where a theory must acquire more precise scope Herzog et al. (2015) Harrison & Bex (2015).

Publication selection and researcher flexibility can exaggerate estimates when many combinations of position, color, size, flanker identity, and exclusion thresholds are available. Preregistration, accessible raw trials, stimulus code, preregistered criteria, and independent replication are valuable safeguards. Meta-analytic pooling requires harmonized outcome measures and explicit handling of heterogeneity rather than combining incomparable crowding thresholds. This monograph follows a critical narrative method, not an exhaustive systematic-review protocol; its conclusions are correspondingly organized by designs and competing explanations instead of purported pooled universal effect sizes Strasburger (2020) Schwetlick & Herzog (2025).

Individual Differences and the Geography of Recognition

Peripheral crowding varies among observers and within visual-field locations. Individual-level maps may be more informative than group averages for understanding reading, visual search, and interface difficulty. Variation may reflect cortical geometry, spatial precision, fixation behavior, perceptual experience, or their combination. A useful model predicts an individual's threshold function across locations and tasks, then evaluates its stability across sessions. Simple rankings based on one brief test risk conflating enduring perceptual organization with measurement noise or task engagement Greenwood et al. (2017) Kurzawski et al. (2025).

The 2025 study by Kurzawski and colleagues makes the neural correlates of this variation especially concrete. With paired psychophysical and fMRI measurements in 49 observers, the investigators linked the capacity for uncrowded letters to visual area V4 surface size, with a reported twofold range in crowding distance across healthy adults. This finding shows how variability can help discriminate cortical-map hypotheses; it does not assign an individual's recognition capacity to a single anatomical determinant. Retinotopic maps, uncertainty estimates, and measurement definitions form part of the claimed result and should accompany its interpretation Kurzawski et al. (2025).

Across developmental groups, visual disorders, and differing experimental conventions, the same word crowding can refer to different mixtures of mechanisms. Measurement invariance is therefore a substantive scientific problem. Testing the same stimulus geometry, criterion, and response set across observers may require accommodations for acuity and comprehension, which can themselves change the task. Hierarchical modeling can estimate individual variation while retaining group-level uncertainty. Longitudinal research is particularly valuable for separating persistent perceptual characteristics from learning, aging, and changing fixation strategy Greenwood et al. (2026) Chung (2007).

Recent Research: Cortical Scaling and Contextual Computation

Contemporary research has strengthened the connection between behavioral crowding and measured neural organization while preserving theoretical plurality. The 2025 V4 study is important because it ties a defined psychophysical quantity to specific retinotopic-map measurements and estimates a proposed conserved cortical distance. Its reported association explains a substantial but incomplete proportion of individual variability, so residual differences remain scientifically informative. Neuroimaging does not by itself distinguish pooling, feedback segmentation, or other computations implemented within the measured territory Kurzawski et al. (2025).

Other recent work emphasizes contextual integration. Cicchini, D'Errico, and Burr formulated a model in which target and surrounding signals are combined according to information reliability. This framing treats the seemingly harmful effect of flankers as a consequence of useful contextual computation under other task demands. It creates predictions for the direction and magnitude of errors, especially when stimulus uncertainty is deliberately manipulated. A good model must remain falsifiable under changed priors, flanker identities, and target reliability, rather than merely interpreting an error after it occurs Cicchini et al. (2022).

The 2025 review by Schwetlick and Herzog situates spatial interactions within perceptual grouping, object organization, and unresolved mechanistic questions. Its significance lies in reminding the field that a single local spacing law is one layer of explanation, not an exhaustive model of visual appearance. Evidence from human behavior, physiology, and computational reconstruction should be compared by matched predictions and levels of analysis. A prospective research program combines global-layout manipulations with quantitative retinal geometry and noninvasive neural measures in the same observers Schwetlick & Herzog (2025) Herzog et al. (2015).

Recent Research: Developmental and Foveal Refinement

Newer developmental work shows how crowding should be examined outside a framework restricted to adult peripheral vision. Greenwood, Kyprianou, and Dekker report that foveal acuity and flanker-induced interference do not mature at identical rates. The final sample of 89 children and 30 adults performed an orientation task with explicit isolated and flanked conditions. This separation is vital: it prevents lower accuracy in younger participants from automatically being attributed to immaturity in target segregation when the isolated target may also be harder to resolve Greenwood et al. (2026).

The study's systematic review of earlier developmental datasets also illustrates the challenge of harmonizing psychophysical measures. Different target classes, contrast polarities, center-to-center versus edge-to-edge spacing definitions, and threshold criteria produce apparent differences in developmental timing. The result strengthens the case for standardized geometry and continuous measurement across age. It does not prove that all foveal and peripheral crowding shares one cortical mechanism or that a single age cutoff can characterize typical development. Such claims require independent neural or intervention-based convergence Greenwood et al. (2026) Strasburger (2020).

Recent accounts of foveal and extrafoveal interactions continue to ask whether spatial scaling changes continuously across the center of gaze. Small eccentricities demand careful fixation monitoring because a few minutes of visual angle can alter the proposed underlying condition. Well-designed tests should vary eccentricity densely near the center while independently measuring unflanked resolution and flanker-dependent critical separation. These methods can reveal whether a simple proportional law extrapolates to the foveola or whether distinct parameter regimes are needed Greenwood et al. (2026) Schwetlick & Herzog (2025).

Human Factors, Artificial Vision, and the Artificial Era

Human interfaces are frequently designed as if everything visible on a display is equally available to recognition. Peripheral vision research offers a different engineering starting point: model what information a user can extract from the current fixation, then assess how display density and grouping affect the next decision. Critical warnings, crowded dashboards, navigation controls, and augmented overlays can be evaluated using eye tracking and objective task performance. The design goal is to support accurate recognition and useful gaze transitions in the actual visual environment, rather than simply satisfy an isolated symbol-size standard Rosenholtz (2016) Ringer et al. (2021).

Computer vision and artificial perceptual systems provide opportunities for comparative modeling. A model with limited receptive fields or peripheral statistical compression may reproduce some human confusions and facilitate efficient search; performance resemblance alone does not establish that it shares human neural organization. A rigorous cross-system analysis defines comparable inputs, tasks, resource constraints, loss functions, and output error distributions. Such experiments can separate benefits of aggregate coding from costs to individual-object identification and test whether recurrent grouping or selective resolution yields robust generalization across visual scenes Freeman & Simoncelli (2011) Balas et al. (2009).

Within the Artificial Era, the scientific importance of crowding extends to relations between different systems of perceptual access. Human interfaces may increasingly be selected, summarized, or generated by artificial agents whose visual representations and failure modes differ from human peripheral encoding. The comparison calls for explicit operational tests rather than transferring human psychological terminology by analogy. The empirical foundation remains human crowding research; proposed cross-system principles become research hypotheses only when a computational system's observables and evaluation criteria are defined. This distinction preserves the scientific value of both human visual psychophysics and artificial perception Rosenholtz et al. (2012) Schwetlick & Herzog (2025).

Conclusion

Peripheral vision supports extended environmental sampling through representations that preserve useful information while limiting the recognition of nearby individual objects. Crowding is one of the most consequential measurable expressions of this organization. Historical letter experiments, psychometric scaling, modern grouping manipulations, and quantitative models establish that recognition is influenced by target–flanker spacing, eccentricity, feature similarity, position, and larger-scale object structure. Its causes are distributed across encoding, integration, segmentation, and access to object identity, with the contribution of each depending on the experimental demand Bouma (1970) Pelli & Tillman (2008) Schwetlick & Herzog (2025).

The empirical record supports several complementary constraints without resolving them into a single universally accepted mechanism. Spatial-pooling and summary-statistic models explain systematic information loss; positional averaging accounts for particular errors; recurrent grouping explains release under changed configurations; attention and eye movements shape how information becomes available. The 2025 association with V4 surface area, the developmental evidence on childhood foveal crowding, and controlled scene experiments connect classical psychophysics to cortical, developmental, and naturalistic questions. Causal localization, prediction across stimulus classes, and longitudinal generalization remain active problems Kurzawski et al. (2025) Greenwood et al. (2026) Ringer et al. (2021).

The resulting scientific synthesis identifies a productive research direction: measure visual information and errors at each step from calibrated stimulation through retinal position, perceptual organization, target selection, and task performance. Models should be evaluated against full response distributions, held-out grouping configurations, individual spatial maps, eye-movement behavior, and ecologically relevant outcomes. The field's conceptual achievement is an operational account of why an item can be plainly present in the visual field while its identity is unavailable amid neighbors. That problem remains central to the psychology of recognition and the design of systems intended for human perception Whitney & Levi (2011) Rosenholtz (2016).

References

  • Balas, B., Nakano, L., & Rosenholtz, R. (2009). A summary-statistic representation in peripheral vision explains visual crowding. Journal of Vision, 9(12), Article 13. https://doi.org/10.1167/9.12.13
  • Bouma, H. (1970). Interaction effects in parafoveal letter recognition. Nature, 226, 177–178. https://doi.org/10.1038/226177a0
  • Chung, S. T. L., Mansfield, J. S., & Legge, G. E. (1998). Psychophysics of reading. XVIII. The effect of print size on reading speed in normal peripheral vision. Vision Research, 38(19), 2949–2962. https://doi.org/10.1016/S0042-6989(98)00072-8
  • Chung, S. T. L., Levi, D. M., & Legge, G. E. (2001). Spatial-frequency and contrast properties of crowding. Vision Research, 41(14), 1833–1850. https://doi.org/10.1016/S0042-6989(01)00071-2
  • Chung, S. T. L. (2007). Learning to identify crowded letters: Does it improve reading speed? Vision Research, 47(25), 3150–3159. https://doi.org/10.1016/j.visres.2007.08.017
  • Cicchini, G. M., D’Errico, G., & Burr, D. C. (2022). Crowding results from optimal integration of visual targets with contextual information. Nature Communications, 13, Article 5741. https://doi.org/10.1038/s41467-022-33508-1
  • Dakin, S. C., Bex, P. J., Cass, J. R., & Watt, R. J. (2009). Dissociable effects of attention and crowding on orientation averaging. Journal of Vision, 9(11), Article 28. https://doi.org/10.1167/9.11.28
  • Francis, G., Manassi, M., & Herzog, M. H. (2017). Neural dynamics of grouping and segmentation explain properties of visual crowding. Psychological Review, 124(4), 483–504. https://doi.org/10.1037/rev0000070
  • Freeman, J., & Simoncelli, E. P. (2011). Metamers of the ventral stream. Nature Neuroscience, 14, 1195–1201. https://doi.org/10.1038/nn.2889
  • Greenwood, J. A., Bex, P. J., & Dakin, S. C. (2009). Positional averaging explains crowding with letter-like stimuli. Proceedings of the National Academy of Sciences, 106(31), 13130–13135. https://doi.org/10.1073/pnas.0901352106
  • Greenwood, J. A., Szinte, M., Sayim, B., & Cavanagh, P. (2017). Variations in crowding, saccadic precision, and spatial localization reveal the shared topology of spatial vision. Proceedings of the National Academy of Sciences, 114(17), E3573–E3582. https://doi.org/10.1073/pnas.1615504114
  • Greenwood, J. A., Kyprianou, M., & Dekker, T. M. (2026). The development of visual acuity and crowding reveals the slow fine-tuning of foveal vision. Scientific Reports, 16, Article 3234. https://doi.org/10.1038/s41598-025-33169-2
  • Harrison, W. J., & Bex, P. J. (2015). A unifying model of orientation crowding in peripheral vision. Current Biology, 25(24), 3213–3219. https://doi.org/10.1016/j.cub.2015.10.052
  • He, S., Cavanagh, P., & Intriligator, J. (1996). Attentional resolution and the locus of visual awareness. Nature, 383, 334–337. https://doi.org/10.1038/383334a0
  • Herzog, M. H., Sayim, B., Chicherov, V., & Manassi, M. (2015). Crowding, grouping, and object recognition: A matter of appearance. Journal of Vision, 15(6), Article 5. https://doi.org/10.1167/15.6.5
  • Kurzawski, J. W., Qiu, B. S., Majaj, N. J., Benson, N. C., Pelli, D. G., & Winawer, J. (2025). Human V4 size predicts crowding distance. Nature Communications, 16, Article 3876. https://doi.org/10.1038/s41467-025-59101-w
  • Levi, D. M. (2008). Crowding—an essential bottleneck for object recognition: A mini-review. Vision Research, 48(5), 635–654. https://doi.org/10.1016/j.visres.2007.12.009
  • Manassi, M., Sayim, B., & Herzog, M. H. (2012). Grouping, pooling, and when bigger is better in visual crowding. Journal of Vision, 12(10), Article 13. https://doi.org/10.1167/12.10.13
  • Nandy, A. S., & Tjan, B. S. (2012). Saccade-confounded image statistics explain visual crowding. Nature Neuroscience, 15, 463–469. https://doi.org/10.1038/nn.3021
  • Pelli, D. G., Palomares, M., & Majaj, N. J. (2004). Crowding is unlike ordinary masking: Distinguishing feature integration from detection. Journal of Vision, 4(12), 1136–1169. https://doi.org/10.1167/4.12.12
  • Pelli, D. G., & Tillman, K. A. (2008). The uncrowded window of object recognition. Nature Neuroscience, 11, 1129–1135. https://doi.org/10.1038/nn.2187
  • Ringer, R. V., Coy, A. M., Larson, A. M., & Loschky, L. C. (2021). Investigating visual crowding of objects in complex real-world scenes. i-Perception, 12(2), 2041669521994150. https://doi.org/10.1177/2041669521994150
  • Rosenholtz, R., Huang, J., & Ehinger, K. A. (2012). Rethinking the role of top-down attention in vision: Effects attributable to a lossy representation in peripheral vision. Frontiers in Psychology, 3, Article 13. https://doi.org/10.3389/fpsyg.2012.00013
  • Rosenholtz, R. (2016). Capabilities and limitations of peripheral vision. Annual Review of Vision Science, 2, 437–457. https://doi.org/10.1146/annurev-vision-082114-035733
  • Schwetlick, L., & Herzog, M. H. (2025). Visual crowding. Annual Review of Vision Science, 11, 359–383. https://doi.org/10.1146/annurev-vision-110423-024409
  • Strasburger, H., Rentschler, I., & Jüttner, M. (2011). Peripheral vision and pattern recognition: A review. Journal of Vision, 11(5), Article 13. https://doi.org/10.1167/11.5.13
  • Strasburger, H. (2020). Seven myths on crowding and peripheral vision. i-Perception, 11(3), 2041669520913052. https://doi.org/10.1177/2041669520913052
  • Toet, A., & Levi, D. M. (1992). The two-dimensional shape of spatial interaction zones in the parafovea. Vision Research, 32(7), 1349–1357. https://doi.org/10.1016/0042-6989(92)90227-A
  • van den Berg, R., Roerdink, J. B. T. M., & Cornelissen, F. W. (2010). A neurophysiologically plausible population code model for feature integration explains visual crowding. PLoS Computational Biology, 6(1), e1000646. https://doi.org/10.1371/journal.pcbi.1000646
  • Whitney, D., & Levi, D. M. (2011). Visual crowding: A fundamental limit on conscious perception and object recognition. Trends in Cognitive Sciences, 15(4), 160–168. https://doi.org/10.1016/j.tics.2011.02.005