Parallax Imaging Techniques

• Parallax imaging is a set of methods that convert changes in viewpoint or observation geometry into measurable displacements to derive depth, distance, and phase information.
• It employs geometric principles such as baseline disparity, optical flow, and epipolar constraints to facilitate precise reconstruction in applications ranging from astrometry to computational imaging.
• Applications include astronomical astrometry, microlensing mass determination, super-resolution in satellite imaging, and correction of detector-induced shifts in diffraction techniques.

Parallax imaging denotes a family of imaging and inference methods in which a change of viewpoint, propagation geometry, or effective observation geometry is converted into a measurable displacement and then into distance, depth, phase, motion, or physically consistent reprojection. In its classical astronomical form, parallax is the apparent change in direction to an object when the observer changes position; in modern usage the same geometric idea appears as stereo disparity, microlensing parallax, epipolar displacement, detector-depth-induced image shift, and aberration-induced virtual-view diversity in phase retrieval (Pössel, 2017, Lee, 2017, Wang et al., 2019, Kuster et al., 2024, Varnavides et al., 24 Jul 2025).

1. Geometric principle and common mathematical structure

The most basic parallax-imaging model uses two observation points separated by a known baseline $b$, an unknown distance $d$, and a parallax angle $\theta$. In the small-angle regime, the distance is approximated by

$d \approx \frac{b}{\theta},$

while the exact classroom-scale formulation can be written with the law of sines for the triangle formed by the two observer positions and the target (Pössel, 2017). The same structure reappears in stereo imaging, where a 3D point projects to $(u_L,v)$ and $(u_R,v)$ in rectified left and right images, producing disparity

$d = u_L - u_R,$

with depth satisfying $Z \propto fB/d$ for focal length $f$ and baseline $B$ (Wang et al., 2019).

Remote-sensing and burst-imaging formulations preserve the same baseline–shift logic but express it in image motion rather than angular coordinates. For SkySat push-frame imagery, elevation changes $d$0 induce a frame-to-frame shift $d$1 that scales as

$d$2

with satellite baseline $d$3 and altitude $d$4 (Anger et al., 2021). In microlensing, the “viewpoint change” is not a rigid image disparity but the change in observer trajectory: annual Earth motion or simultaneous space–ground observations perturb the standard Paczyński light curve, and the resulting microlens parallax is encoded in

$d$5

so that geometric observables determine the lens mass when combined with an Einstein-angle measurement (Lee, 2017).

A distinct but related contrast appears in perspective imaging and stitching. Under perspective projection, depth-dependent displacement prevents a single 2D warp from aligning all scene points. Orthographic projection is parallax-free because scaling does not depend on depth $d$6; this difference underlies methods that convert perspective measurements into orthographic or 3D-consistent renderings before stitching (Fotouhi et al., 2020). Across these domains, the shared structure is that parallax is not merely an artifact: it is a measurable encoding of scene geometry or system geometry.

2. Astrometric parallax in astronomy

Classical astronomical parallax measures the apparent shift of a nearby star relative to distant background stars as Earth moves along its orbit, typically using a $d$7 six-month baseline (Pössel, 2017). In catalog-scale astrometry, repeated imaging over years is modeled as position plus proper motion plus annual parallax. The URAT Parallax Catalog expresses this in the standard linear form

$d$8

with parallax factors computed from the JPL DE405 ephemeris and Green’s formulation. URAT used Northern Hemisphere exposures from April 2012 to June 2015, published 112,177 parallaxes, and reported average parallax precision of $d$9 mas for stars having no known parallax and $\theta$0 mas for stars matched to external parallax sources (Finch et al., 2016).

A more specialized astrometric realization is HST/WFC3 spatial scanning, where the telescope slews during the exposure so that each star forms a long trail rather than a point source. This converts a single centroid estimate into thousands of cross-scan minirow measurements, increasing source sampling by roughly a thousand-fold and enabling changes in source positions to be measured to $\theta$1–$\theta$2 microarcseconds. The method was designed to extend HST parallax work to bright Cepheids at distances up to $\theta$3 kpc and was demonstrated on SY Aurigae using five epochs separated by six months (Riess et al., 2014). The same scanning mode also provides photometry of bright Milky Way Cepheids on the same flux scale as extragalactic Cepheids, which is relevant because the distance-scale application requires astrometry and photometry on a common instrumental system (Riess et al., 2014).

These astrometric implementations clarify a central distinction. In catalog parallax work, parallax imaging is repeated precision imaging of stellar positions. In spatial-scanning work, the detector is repurposed so that the scan itself becomes the precision astrometric observable. The underlying quantity remains the same trigonometric parallax, but the imaging strategy changes the attainable precision regime.

3. Microlensing, high-resolution imaging, and astrometric reconstruction

In microlensing, parallax imaging denotes a geometric reconstruction problem rather than a direct positional shift on a static image. A lens of mass $\theta$4 at distance $\theta$5 and a source at distance $\theta$6 define an Einstein angle

$\theta$7

and a standard point-lens light curve yields $\theta$8, $\theta$9, and $d \approx \frac{b}{\theta},$0, but leaves a degeneracy in $d \approx \frac{b}{\theta},$1, $d \approx \frac{b}{\theta},$2, $d \approx \frac{b}{\theta},$3, and $d \approx \frac{b}{\theta},$4 (Lee, 2017). Microlensing parallax breaks this degeneracy by exploiting annual observer acceleration, space–ground baselines, or terrestrial baselines, all of which produce measurable deviations from the symmetric Paczyński curve. Combined with $d \approx \frac{b}{\theta},$5, the parallax vector $d \approx \frac{b}{\theta},$6 yields the lens mass through $d \approx \frac{b}{\theta},$7 (Lee, 2017).

High-resolution imaging contributes two additional degeneracy breakers: measurement of the lens–source relative proper motion and measurement of the lens flux. If the lens and source can later be resolved, then

$d \approx \frac{b}{\theta},$8

and the resulting $d \approx \frac{b}{\theta},$9 can be combined with $(u_L,v)$0 for a geometric mass determination (Lee, 2017). If the lens flux $(u_L,v)$1 is measured and the lens is assumed to be a normal star, stellar isochrones yield a mass estimate that can be combined with the microlensing mass ratio $(u_L,v)$2. This strategy is exemplified by OGLE‑2014‑BLG‑0124, where simultaneous ground–Spitzer observations produced an excellent parallax measurement and Keck II adaptive-optics imaging measured the lens flux. The combined constraints gave a host-star mass $(u_L,v)$3, lens distance $(u_L,v)$4 kpc, and planet mass $(u_L,v)$5 at $(u_L,v)$6 AU (Beaulieu et al., 2017).

Astrometric microlensing extends this geometric reconstruction into the image-centroid domain. For a point lens, the unresolved-image centroid shift is

$(u_L,v)$7

so a time series of centroid positions traces an ellipse whose scale directly yields $(u_L,v)$8 (Lee, 2017). This suggests a broader use of “parallax imaging” in microlensing: not only light-curve asymmetries, but also resolved imaging, centroid motion, and lens-flux measurements become complementary geometric observables of the same event.

4. Multi-view reconstruction, super-resolution, and large-baseline synthesis

In computational imaging, parallax imaging commonly refers to exploiting disparity across views to reconstruct higher-resolution or novel-view content. Stereo image super-resolution uses a low-resolution stereo pair and the fact that corresponding features lie on the same epipolar row but at different horizontal positions. PASSRnet formalizes this with a parallax-attention mechanism that computes row-wise correspondences across the full epipolar line, avoiding a fixed disparity range and using valid masks, photometric consistency, smoothness, and cycle consistency to handle occlusion and large disparity variation (Wang et al., 2019). The same paper introduced Flickr1024 as a large stereo SR dataset and reported state-of-the-art performance on Middlebury, KITTI 2012, and KITTI 2015 (Wang et al., 2019).

For push-frame satellite bursts, parallax becomes relief-dependent apparent motion. SkySat imagery is modeled with a Plane+Parallax decomposition in which a global affine transform $(u_L,v)$9 stabilizes a reference plane and a dense field $(u_R,v)$0 captures residual parallax: $(u_R,v)$1 A two-stage pipeline first factorizes pairwise flows into affine motion plus coarse parallax and then estimates a single multi-frame parallax field by robust optical flow, using brightness constancy, gradient constancy, and smoothness terms (Anger et al., 2021). The estimated field improves multi-frame super-resolution in scenes with elevation changes and also provides a coarse 3D surface model (Anger et al., 2021).

Light-field view synthesis extends the same idea to much larger angular extrapolation. SLSC synthesizes views far outside the captured angular baseline by quantizing disparity into stratified layers, warping each layer separately, compositing with a near-over-far rule, and then using a GAN for parallax correction and occlusion completion (Chen et al., 2019). The method explicitly targets large perspective shifts and reports more than $(u_R,v)$2 dB improvement over prior light-field synthesis algorithms, with results shown for baseline extension ratios up to $(u_R,v)$3 (Chen et al., 2019).

A conceptually different realization is binocular parallax stereo imaging with pseudo-thermal speckle illumination. Here correspondence is not driven by static texture but by temporal intensity-fluctuation correlation between two synchronized cameras. The paper reports one-pixel authentic matching precision and emphasizes that the method performs well when the object’s superficial characteristics are not obvious, for example when its surface reflectivity is constant (Zhu et al., 2014). This is still parallax imaging, but the discriminative signal resides in second-order fluctuation statistics rather than conventional appearance.

5. Parallax-robust stitching and reprojection

Image stitching is one of the domains where parallax is most visibly problematic, because depth variation makes global homographies invalid. In medical X-ray imaging, this issue is addressed by abandoning purely 2D mosaicing. A perspective C-arm acquisition is first backprojected into a 3D volume

$(u_R,v)$4

its 3D Fourier transform is computed, a central slice corresponding to an orthographic projection plane is extracted, and the stitched image is obtained by inverse 2D Fourier transform (Fotouhi et al., 2020). Because the final image is an orthographic projection of a common 3D representation, it is parallax-free by construction. The resulting system achieved SSIM $(u_R,v)$5 and PSNR $(u_R,v)$6 dB on test data and produced stitched images suitable for metric measurements directly on the 2D plane (Fotouhi et al., 2020).

A geometry-driven alternative in perspective image stitching is the epipolar displacement field. Instead of allowing arbitrary 2D nonrigid deformation, the method starts from the infinite homography

$(u_R,v)$7

and models residual parallax as displacement constrained to epipolar lines, with the displacement field represented by thin-plate splines (Yu et al., 2023). Pixels are inversely warped according to this epipolar displacement field, which improves alignment while preserving panorama projectivity more effectively than unconstrained elastic warps. The paper reports smaller average distances to epipolar lines in non-overlap regions than robust elastic warping, indicating better preservation of global projective structure (Yu et al., 2023).

For very large parallax, the 2D-warp paradigm is replaced altogether by 3D reconstruction and reprojection. PIS3R uses a Visual Geometry Grounded Transformer to estimate intrinsics, extrinsics, and dense point maps, reprojects the reconstructed point cloud to a designated reference view, and then refines the initial stitching with a point-conditioned image diffusion module (Zhu et al., 6 Aug 2025). Quantitatively, it reported PSNR $(u_R,v)$8, SSIM $(u_R,v)$9, LPIPS $d = u_L - u_R,$0, and a Sampson epipolar error of $d = u_L - u_R,$1 px, with registration success rate $d = u_L - u_R,$2 on synthetic data and $d = u_L - u_R,$3 on real data (Zhu et al., 6 Aug 2025). A plausible implication is that, once parallax becomes very large, “stitching” is better understood as 3D reconstruction plus rendering than as 2D alignment.

6. Detector, diffraction, and phase-retrieval manifestations

In some systems, parallax does not arise from a macroscopic baseline between cameras but from propagation inside the detector or the imaging kernel itself. Thick silicon sensors in coherent diffraction imaging provide a clear example. For a $d = u_L - u_R,$4 thick fully depleted sensor with $d = u_L - u_R,$5 or $d = u_L - u_R,$6 pixels, a photon entering at angle $d = u_L - u_R,$7 is absorbed at a random depth and its charge cloud drifts to the readout plane. The measured point of detection is therefore shifted laterally relative to the point of incidence, with average displacement scaling as

$d = u_L - u_R,$8

The resulting PSF becomes asymmetric and drop-like at high angles, per-pixel intensity falls to less than $d = u_L - u_R,$9 at $Z \propto fB/d$0 for 12 keV photons, and for a planar detector with $Z \propto fB/d$1 pixels covering $Z \propto fB/d$2 at 12 keV the effective sample resolution is degraded by a factor of approximately $Z \propto fB/d$3 (Kuster et al., 2024). Here “parallax imaging” denotes a position-dependent detector blur kernel rather than a scene-level disparity.

Angular-sensitive powder diffraction tomography exhibits another geometric version. A diffracting point offset by $Z \propto fB/d$4 from the rotation axis produces a detector shift

$Z \propto fB/d$5

which appears as an apparent Bragg-angle offset

$Z \propto fB/d$6

The paper shows that this parallax contribution is additive with other angular offsets in the first-moment sinogram and that, for full $Z \propto fB/d$7 scans, parallax has no impact on reconstructions of angular information because the sinusoidal contribution averages to zero over a full rotation (Modregger et al., 2024). This is a particularly explicit case in which parallax is a correctable geometric bias rather than a desired signal.

In direct ptychography, parallax appears in yet another sense: aberrations generate virtual view diversity across bright-field detector pixels. The paper formulates parallax imaging as a quadratic approximation to the direct ptychography kernel and defines a vBF diversity ratio

$Z \propto fB/d$8

with the practical criterion $Z \propto fB/d$9 for successful scan upsampling (Varnavides et al., 24 Jul 2025). Under sufficient defocus, even $f$0 scan upsampling recovers a direct-ptychography transfer function close to the analytical ideal, while at lower defocus only $f$1 upsampling remains reliable (Varnavides et al., 24 Jul 2025). This suggests that, in wave-optical imaging, parallax can be interpreted as controlled kernel diversity that relaxes sampling constraints rather than as an external viewpoint change.

Across these detector and diffraction settings, the term “parallax imaging” is broadened but not emptied. The common element is still geometry-dependent displacement carrying recoverable information; what changes is whether the geometry is that of an observer baseline, a lensing trajectory, a detector volume, or an aberrated wavefront.