Two well-documented biases in early vision have so far been studied in isolation. The first is a bias toward darkness. Natural scenes contain more and stronger dark features than bright ones, and the retina and cortex amplify this imbalance, so a dark bar drives visual cortex harder than an equally strong bright bar. The second is a bias toward the body's straight-ahead direction. A stimulus at a fixed spot on the retina evokes a larger cortical response, and faster eye and hand responses, when that spot happens to line up with the direction the body faces. All the evidence for the straight-ahead bias comes from black-and-white checkerboards. Those patterns drive the "darker-than-background" and "brighter-than-background" pathways at once and average away any difference between them. So nobody knows whether the straight-ahead boost is a uniform volume knob on all visual input, or whether it selectively strengthens one contrast polarity.
We propose a human EEG and psychophysics study that answers this directly. We measure full contrast-response curves for pure dark and pure bright probes at matched retinal locations that are either body-straight-ahead or body-eccentric. A second experiment rotates the head on the trunk and uses contrast polarity as a tag to find which body axis anchors the boost. The core claim is testable and either outcome is informative. A pure multiplicative gain would predict that the dark-to-bright response ratio is identical at straight-ahead and eccentric positions. Any reliable polarity-by-position interaction would show that the egocentric prior acts selectively on the channels that carry the most information in natural scenes. An interaction would also show where in the response nonlinearity the boost operates.
Background and what is missing
The dark side of the story supplies the motivation. Predicting Cortical Dark/Bright Asymmetries from Natural Image Statistics and Early Visual Transforms filtered calibrated natural images into dark and bright components. The paper reports that dark amplitude exceeds bright by a factor of 1.4, that the bias grows with contrast and is strongest at low spatial frequencies, and that simulated retinal output to cortex is 1.9 times larger for a dark bar. Its model predicts a roughly 40% lower discrimination threshold for darks. Predicting and Manipulating Cone Responses to Naturalistic Inputs found the asymmetry already present in primate cones: they adapt to increments in about 23 ms but recover from decrements in about 94 ms, and on brighter backgrounds they become more sensitive to decrements. That paper also shows linear and linear-nonlinear models fail to capture these dynamics. This constrains our analysis: we cannot assume a symmetric front end and then add the straight-ahead effect on top.
The straight-ahead side supplies the paradigm. Dynamics of the straight-ahead preference in human visual cortex kept head and chest fixed and shifted gaze 10° left or right, so that the same retinal location fell either on or off the body midline. With 64-channel EEG in 58 people, it found a straight-ahead enhancement of the contralateral P1 at about 70 ms, no effect on the early feedforward C1, an enhancement of the later C2 at 130–160 ms, and lower pre-stimulus alpha power over the relevant hemisphere. An egocentric straight-ahead bias in primate's vision reviews macaque single units, human fMRI and behavior. It reports up to 25% larger V1/V2 responses and roughly 10 ms faster reactive eye and hand responses, and interprets the bias as a history-driven, ecologically shaped form of attention. Both works used contrast-symmetric checkerboards. Both kept head and trunk aligned, so whether the bias is anchored to the nose or the sternum is unknown. The EEG study saw no reaction-time effect, which its authors attribute to a 150 cm viewing distance.
The broader reference-frame literature constrains interpretation. Maps of Visual Space in Human Occipital Cortex Are Retinotopic, Not Spatiotopic showed that all twelve tested human visual areas map space in eye-centered coordinates. It also showed that apparent world-centered voxels were a low signal-to-noise artifact that failed cross-validation. We therefore treat the straight-ahead effect as a gain modulation on a retinotopic map rather than a remapping, and we adopt that paper's insistence on cross-validated, held-out tests of any frame claim. Eye-Centered Representation of Optic Flow Tuning in the Ventral Intraparietal Area shows that one brain area can hold visual signals in eye coordinates and vestibular signals in body coordinates. Its stimuli were bright dots only. This is the model for our second experiment, which decouples head from trunk with a displacement-style index.
Why might the boost be polarity-selective at all? Cottereau and colleagues argue the bias is history-driven, meaning it reflects where behaviorally relevant input has tended to fall. Natural input is dark-dominated, and dark dominance is strongest at high contrast. If the straight-ahead gain is tuned by experience, the most plausible place for it to act is the high-contrast, dark-carrying end of the response curve. That predicts a larger straight-ahead boost for darks that appears as a change in maximum response rather than in sensitivity. The opposite outcome is also coherent: darks are already amplified and closer to saturation, so a boost could be largest for brights. We do not need to know the sign in advance. We need a design that distinguishes multiplicative from selective gain cleanly.
Core Claim
The straight-ahead preference is not a polarity-blind gain. We claim that the ratio of dark to bright responses in early visual cortex changes depending on whether a fixed retinal location is body-straight-ahead. We further claim that where this change occurs on the contrast-response curve tells us whether the egocentric prior acts on contrast sensitivity or on response amplitude. As a secondary claim, using polarity as a tag lets us determine whether the prior is anchored to the head or the trunk, a question the existing paradigm cannot answer.
The test is stated compactly as follows. For each polarity $p$ (dark or bright) and body position $b$ (straight-ahead or eccentric), we fit a saturating contrast-response function to evoked amplitude $R$ at probe contrast $c$:
$$R_{p,b}(c) = R^{max}_{p,b}\,\frac{c^n}{c^n + (c^{50}_{p,b})^n} + R_0 .$$
Here $R^{max}$ is the response ceiling, $c^{50}$ is the contrast giving half the ceiling, $n$ is a shared steepness exponent, and $R_0$ is baseline. The pure multiplicative-gain hypothesis implied by prior work holds that $R^{max}_{p,\text{SA}} / R^{max}_{p,\text{ecc}}$ is the same for both polarities and $c^{50}$ is unchanged. Our hypothesis is that one of these ratios differs between polarities.
Research Program
The first component is a short image-statistics analysis. It gives the polarity hypothesis a prior and a direction rather than leaving it a guess. We will apply the dark/bright decomposition of Cooper and Norcia to the van Hateren database (200 calibrated images) and the McGill calibrated color database (80 images). Instead of pooling across each image, we compute the dark-to-bright amplitude ratio as a function of image position, separately for the central horizontal band and the 10° flanks, at the spatial frequency of our probes.
Camera-centered images are a stand-in for head-centered input, not trunk-centered input. We state this limitation openly. The analysis cannot settle the question. It only predicts whether the straight-ahead region is unusually dark-rich, which would support a dark-favoring boost. Cooper and Norcia note that their ratios depend on how local contrast is normalized, so we report results under two normalization pool sizes. We draw directional conclusions only where they agree.
The second component is Experiment 1, a polarity version of the Bogdanova gaze-shift paradigm. Head and chest stay aligned with the screen center. Gaze is held at 10° left or right. Brief probes are flashed at 10° retinal eccentricity, so that on half the trials they fall on the body midline and on half at 20° body eccentricity, with retinal position identical. The probes are uniform low-spatial-frequency blobs or bars that are either darker or brighter than a mid-gray background, at five Weber contrasts spanning threshold to near-saturation. We also include the original contrast-symmetric checkerboard as an anchor condition.
We record 64-channel EEG and extract C1, contralateral P1, and C2 by component latency and topography following Bogdanova and colleagues. In interleaved blocks we measure detection thresholds by staircase, and saccadic and manual reaction times. Following Bogdanova's own explanation for their null behavioral result, the behavioral blocks are run at a 45 cm peripersonal viewing distance as well as the original distance.
Darkening a region lowers mean luminance and can shift pupil size and global adaptation. To control for this, each probe is embedded in a fixed annulus whose luminance compensates, so that space-averaged luminance is constant across polarities. We also track pupil diameter as a covariate. Background luminance is run at two levels. Angueyra and colleagues show that cones favor decrements more on brighter backgrounds, so a polarity effect that grows with background level would point to a retinal contribution. A polarity effect that is independent of background level would point to a cortical feedback origin, which fits the effect appearing at P1 and C2 but not C1.
The third component is Experiment 2, which separates head from trunk. Participants sit with the trunk rotated 20° relative to the screen while the head remains facing the screen center, and in a mirror condition the head is rotated on a screen-facing trunk. Gaze positions are chosen so that the same retinal location is, across conditions, aligned with the head midline only, the trunk midline only, both, or neither. Because Experiment 1 tells us how polarity shapes the straight-ahead response, we can use the polarity-specific signature as a sharper marker of "straight-ahead-ness" than amplitude alone. For each axis we compute a displacement-style index in the spirit of the VIP study: the fraction of the straight-ahead effect that follows that axis when the two axes are separated. Following Gardner and colleagues, the index is estimated on held-out runs to avoid the low signal-to-noise trap.
Baselines
The naive reference point is the null in which the straight-ahead boost does not exist under these stimuli and polarity effects are constant across body positions. Darks should then be more sensitive at both positions, roughly in line with the 40% lower dark threshold predicted by Cooper and Norcia, and P1 and C2 should not differ between straight-ahead and eccentric conditions. We will run it by fitting a model with polarity terms only. If our data fit this model as well as richer ones under cross-validation, the proposal's premise has failed. It would also mean the published straight-ahead effect does not generalize beyond checkerboards, which is itself worth reporting.
The strongest existing competitor is the symmetric multiplicative gain model implicit in Dynamics of the straight-ahead preference in human visual cortex and An egocentric straight-ahead bias in primate's vision. In this model, straight-ahead position scales responses by one factor regardless of input. The review reports up to a 25% BOLD increase in V1/V2 and roughly 10 ms faster reactive responses. For the EEG components we measure, the file provides timing (P1 at 70 ms, C2 at 130–160 ms) but no effect size, so no reported result was found on Lacuna for the size of the P1 straight-ahead effect. We will run this model as one shared gain parameter on both polarities' contrast-response functions, fit to exactly the same trials, with the same number of free parameters per condition as our selective model except for the polarity-specific gain terms.
Beating it means the following. The polarity-specific model must win by cross-validated likelihood on held-out participants. The 95% interval on the difference between dark and bright gain ratios must exclude zero for P1 or C2. The interaction must hold at both background levels. A difference that appears in only one component or only at one contrast does not count.
The key ablation is the checkerboard anchor condition. It must reproduce the published straight-ahead enhancement in our setup. If it does not, our probes and setup are not measuring the same phenomenon and the polarity comparison is uninterpretable. A second ablation removes the luminance-compensating annulus in a subset of blocks to show how much of any polarity effect arises from changes in mean luminance rather than local contrast polarity.
Why not just…?
Why not just multiply the known dark advantage by the known straight-ahead gain and skip the experiment? That is exactly the multiplicative-gain hypothesis, and nothing in the file tests it. Every straight-ahead measurement used stimuli that drive both polarities together. A checkerboard response can therefore look like a single gain even if the two pathways are boosted differently, since the effects add before measurement. The product prediction is the baseline we test, not a substitute for testing it.
Why not just reanalyze the existing checkerboard EEG by separating dark and bright checks? In a flashed checkerboard the dark and bright checks appear at the same moment within each receptive field's neighborhood. At EEG resolution, their contributions to C1, P1 and C2 cannot be separated after the fact. Separating polarities requires stimuli that contain only one polarity, which is what Experiment 1 introduces.
Why not just study the coordinate frame of the light-from-above prior instead, by tilting observers? It is a natural question at this intersection. The file flags it as untested, citing Lighting-from-above prior in biological motion perception and Dynamics of Response to Perceptual Pop-Out Stimuli in Macaque V1, both run with upright observers. We set it aside for two reasons. First, the frame of the lighting prior has a substantial psychophysics history beyond this file, so novelty there is doubtful. Second, the full design needs body tilt rigs. The polarity-by-straight-ahead question is untouched, cheaper, and runs on standard EEG equipment.
Why not just attribute any polarity effect to eye-in-head position or pupil changes rather than body straight-ahead? This is a real threat. Shifting gaze 10° changes eye position, and dark probes can alter pupil size. Our design addresses it three ways. Mean luminance is compensated and pupil diameter is used as a covariate. Retinal position is matched across conditions. In Experiment 2, the same eye-in-head position occurs with straight-ahead and eccentric trunk or head alignments. An effect that follows eye-in-head position rather than either body axis in Experiment 2 would undercut the egocentric interpretation, and we would report it that way.
Decisive checks
The study succeeds if three things hold. The checkerboard anchor replicates the published straight-ahead enhancement at P1 and C2. The polarity-specific gain model beats the symmetric model on held-out participants with a polarity-by-position interaction whose interval excludes zero. The interaction appears in P1 or C2 but not C1, consistent with the feedback timing reported by Bogdanova and colleagues.
Whether the interaction falls on $R^{max}$ or $c^{50}$ tells us whether the egocentric prior scales response amplitude or contrast sensitivity. The image-statistics prediction tells us whether its sign matches the polarity balance of the straight-ahead region in natural images. In Experiment 2, a cross-validated displacement index near 1 for one body axis and near 0 for the other resolves the head-versus-trunk ambiguity. Intermediate values would indicate a mixed frame, as VIP shows across modalities. A clean null on the interaction, given a replicated anchor, is also decisive. It would establish that the straight-ahead prior is a polarity-blind gain, which constrains any account of it as learned from natural input statistics.
References
- Predicting Cortical Dark/Bright Asymmetries from Natural Image Statistics and Early Visual Transforms
- Predicting and Manipulating Cone Responses to Naturalistic Inputs
- Dynamics of the straight-ahead preference in human visual cortex
- An egocentric straight-ahead bias in primate’s vision
- Maps of Visual Space in Human Occipital Cortex Are Retinotopic, Not Spatiotopic
- Eye-Centered Representation of Optic Flow Tuning in the Ventral Intraparietal Area
- Lighting-from-above prior in biological motion perception
- Dynamics of Response to Perceptual Pop-Out Stimuli in Macaque V1