Image Stereocorrelation Methods

• Image stereocorrelation is a technique that establishes correspondences between paired images to accurately recover 3D shapes and material motion.
• It leverages various cost functions, including normalized cross-correlation, block matching, and optimal transport, to achieve precise disparity and sub-pixel estimations.
• The method finds applications in computer vision and experimental mechanics, integrating digital image correlation with thermo-mechanical analysis.

Searching arXiv for recent and relevant papers on image stereocorrelation and related stereocorrelation methods. arxiv_search(query="image stereocorrelation stereo correlation digital image correlation infrared thermography", max_results=10, sort_by="relevance") Image stereocorrelation, in the literature surveyed here, denotes a family of stereo-matching and full-field metrology procedures in which correspondence is established between paired images, random patterns, or multiview observations so that disparity, three-dimensional coordinates, shape, or material-point motion can be recovered. In computer vision, the term covers local correlation costs such as zero-means normalized cross correlation, block-matching by sum of absolute differences, phase-only correlation, and line-by-line formulations based on optimal transport; in experimental mechanics, it includes stereocorrelation-based digital image correlation, phase-based random-pattern analysis, and projection-based coupling to infrared thermography (Chang et al., 2022, Molimard et al., 2013, Tröger et al., 27 Jun 2026).

1. Historical and conceptual scope

Classical digital image correlation appears in the surveyed literature as an initial reference point. In the frequency-based alternative to classical stereocorrelation, a reference image $f(r,s)$ and a deformed image $g(r,s)$ are related by

$g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$

and the integer part of the shift is obtained by maximizing the cross-correlation

$$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$

The same work presents phase analysis of random patterns as an alternative to classical stereocorrelation, with metrological characterization showing a resolution comparable to classical (stereo) correlation technique $(1/100\text{th pixel})$ and a spatial resolution advantage associated with temporal phase stepping and windowed Fourier transform (Molimard et al., 2013).

In stereocorrelation-based digital image correlation, two synchronized cameras recover the three-dimensional motion of surface points in a Lagrangian reference, whereas infrared thermography measures two-dimensional temperature fields in an Eulerian image plane. The projection-based coupling literature therefore treats stereocorrelation not merely as a disparity-estimation device, but as the kinematic backbone of a thermo-mechanically consistent description of the specimen surface, especially for curved surfaces (Tröger et al., 27 Jun 2026).

Within computer vision, stereo matching is described as one of the most important sub fields of computer vision, with the parallax map transformed into a depth map for use in intelligent fields. The same corpus spans embedded-GPU implementations, line-by-line optimal-transport formulations, non-regular sampling sensors, and end-to-end cost-volume networks, indicating that image stereocorrelation is simultaneously an estimation problem, a systems problem, and a representational problem (Wang et al., 2021, Galeotti et al., 2022, Kwon et al., 2022).

A distinct optical realization is binocular parallax stereo imaging based on intensity-fluctuation correlation. There, the one-to-one mapping inherent in speckle-field correlations turns every illuminated point on an object’s surface into a reliable stereo “feature,” and the method is reported to perform better when the object's superficial characteristics are not obvious, for example its surface reflectivity is constant (Zhu et al., 2014).

2. Correlation costs and correspondence models

A canonical local cost in image stereocorrelation is zero-means normalized cross correlation. For a reference image $I_R(x,y)$, a target image $I_T(x,y)$, and disparity $d\in[0,D)$, the ZNCC cost is defined as

$$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$

The same formulation is algebraically rearranged into

$$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$

so that $g(r,s)$0, $g(r,s)$1, and the analogous target-image terms can be pulled out and pre-computed once via integral-image summations (Chang et al., 2022).

A second classical model is windowed block matching. In the stereo reconstruction method for non-regular sampling sensors, the preliminary high-resolution reconstructions $g(r,s)$2 and $g(r,s)$3 are matched by the sum-of-absolute-differences cost

$g(r,s)$4

The best disparity is the minimizer over an allowed search zone, followed by a left–right consistency check that invalidates pixels whose disparities disagree by more than one (Jonscher et al., 2022).

The “Stereo Matching Based on Visual Sensitive Information” line of work augments a census-based gray-level cost with a visual-sensitivity cost. It defines

$g(r,s)$5

$g(r,s)$6

maps both terms into $g(r,s)$7 by exponential functions, and combines them in

$g(r,s)$8

This pixel-wise cost is then aggregated over a dynamic, cross-shaped region whose arms extend under spatial and gray-difference constraints that vary with the local visual-sensitivity direction $g(r,s)$9 (Wang et al., 2021).

A different correspondence model replaces local windows by global transport. In the optimal-transport framework, one scan-line of each image is normalized into mass vectors $g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$0 and $g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$1, the cost matrix is $g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$2, and the classical Kantorovich problem minimizes $g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$3 over transport plans $g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$4. Entropic regularization yields

$g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$5

with Gibbs-kernel solution

$g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$6

Disparity is extracted by the barycentric shift

$g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$7

This replaces local similarity accumulation by a convex line-by-line transport problem (Galeotti et al., 2022).

3. Sub-pixel localization and phase-based stereocorrelation

Sub-pixel estimation is a central theme in several stereocorrelation traditions. In low-resolution thermal stereo, the ST framework (“Subpixel Thermal”) consists of feature extraction via phase congruency, integer-pixel feature matching, and sub-pixel refinement by phase correlation in the feature space. In a setup based on two $g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$8 px FLIR Lepton 2 cameras with baseline $g(r,s)=f(r-\Delta x,s-\Delta y)+b(r,s),$9 mm, a $$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$0 px disparity error induces a depth uncertainty of almost half a meter, which makes sub-pixel accuracy critical (Zoetgnande et al., 2019).

The phase-congruency stage is explicitly designed for thermal images that lack texture and suffer from brightness drifts. In one dimension,

$$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$1

and a more robust form subtracts a noise floor $$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$2, applies weighting $$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$3, and avoids division by zero. In two dimensions, oriented band-pass filters are applied and the maximum moment

$$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$4

is used for feature detection, with strong features defined by $$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$5. On $$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$6 px thermal frames, setting $$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$7 typically returns $$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$8 features per image, compared to ORB’s $$h(\Delta x,\Delta y)=\iint g(a,b)\,f(a-\Delta x,b-\Delta y)\,da\,db.$$9 and SURF’s $(1/100\text{th pixel})$0; under simulated gain/bias changes $(1/100\text{th pixel})$1, phase congruency retains $(1/100\text{th pixel})$2 of its features up to large $(1/100\text{th pixel})$3 shifts, whereas ORB and SURF drops below $(1/100\text{th pixel})$4 (Zoetgnande et al., 2019).

ST performs sub-pixel refinement by normalized cross-power spectrum and phase-only correlation. For sub-images $(1/100\text{th pixel})$5 and $(1/100\text{th pixel})$6,

$(1/100\text{th pixel})$7

and the local peak is modeled by

$(1/100\text{th pixel})$8

The displacement estimate is obtained by least squares with $(1/100\text{th pixel})$9 and SVD. On $I_R(x,y)$0 thermal data, a $I_R(x,y)$1 window is the best trade-off, and with $I_R(x,y)$2 the method achieves $I_R(x,y)$3 of matches within $I_R(x,y)$4 px, $I_R(x,y)$5 within $I_R(x,y)$6 px, $I_R(x,y)$7 within $I_R(x,y)$8 px, and $I_R(x,y)$9 within $I_T(x,y)$0 px (Zoetgnande et al., 2019).

Phase analysis also appears in a frequency-based alternative to classical stereocorrelation. Each Zone Of Interest is tapered by a bi-triangular window and analyzed by a windowed Fourier transform; around each dominant spatial frequency $I_T(x,y)$1, the phase shift satisfies

$I_T(x,y)$2

Collecting $I_T(x,y)$3 such phase differences leads to the least-squares solution

$I_T(x,y)$4

The reported sub-pixel resolution in $I_T(x,y)$5 is on the order of $I_T(x,y)$6–$I_T(x,y)$7 pixel RMS, with bias $I_T(x,y)$8 px for a noise-free ZOI of $I_T(x,y)$9, and overall resolution comparable or better than classical stereocorrelation while avoiding ad hoc peak fitting (Molimard et al., 2013).

In binocular parallax stereo imaging based on intensity-fluctuation correlation, sub-pixel precision is tied to the sharp peak of the weighted correlation

$d\in[0,D)$0

The matching precision is governed by the speckle transverse correlation length $d\in[0,D)$1, which maps to $d\in[0,D)$2 camera pixel, so the method achieves true $d\in[0,D)$3-pixel disparity precision, and sub-pixel by peak-fitting (Zhu et al., 2014).

4. Efficient and learned computational pipelines

A central engineering issue in image stereocorrelation is the tension between accuracy and throughput. The embedded-GPU ZNCC work addresses this by accelerating the heaviest term

$d\in[0,D)$4

through zigzag scanning with register reuse. The image scan is partitioned into stripes of height $d\in[0,D)$5 and width $d\in[0,D)$6; a $d\in[0,D)$7 block of input pixels is loaded into shared memory once; and within each CUDA block, $d\in[0,D)$8 threads compute one disparity each. The “zigzag” ordering reuses partial sums in registers rather than reloading them from memory, cutting global-memory bandwidth by roughly $d\in[0,D)$9. On a Jetson TX2 GPU, the method is reported as $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$0 faster than the traditional image scanning method and $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$1 faster than the latest NCC method; the full pipeline with domain transformation and winner-take-all reaches real-time processing speed of $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$2 fps for $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$3 pixel images with a maximum disparity of $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$4, and the KITTI 2015 D1-all error rate is $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$5, improving census+FastDT $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$6 by $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$7 (Chang et al., 2022).

The same paper compresses the floating-point cost volume for domain transformation by subtracting a nearly-zero-mean offset, scaling by $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$8, and packing the $$C(x,y,d)=\frac{\sum_{(u,v)\in W}\bigl(I_R(x+u,y+v)-\overline{I_R}(x,y)\bigr)\bigl(I_T(x+u-d,y+v)-\overline{I_T}(x-d,y)\bigr)}{\sigma_R(x,y)\,\sigma_T(x-d,y)}.$$9-bit residual into an $$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$0-bit value plus $$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$1-bit position flags. The reported consequence is $$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$2–$$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$3 faster memory passes and overall $$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$4 faster DT than $$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$5-bit float, while integral-image computation of means and variances remains $$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$6 ms/frame independent of window size (Chang et al., 2022).

Optimal transport provides a different route to efficiency. After entropic regularization, the transport plan is solved by Sinkhorn–Knopp iterations,

$$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$7

with complexity per iteration $$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$8 for a single row. The framework explicitly proposes GPU parallelism, truncation of $$C(x,y,d)=\frac{\sum_W I_R I_T-|W|\overline{I_R}\,\overline{I_T}}{\sqrt{\sum_W I_R^2-|W|\overline{I_R}^2}\cdot\sqrt{\sum_W I_T^2-|W|\overline{I_T}^2}},$$9 to a band of width $g(r,s)$00, and multi-scale acceleration, and it extends to images with occluded regions through an unbalanced-mass construction based on shift and rescale (Galeotti et al., 2022).

Learned pipelines depart from explicit matching windows but still retain a correlation structure. Image-Coupled Volume Propagation initializes a $g(r,s)$01D cost volume $g(r,s)$02 by group-wise correlation,

$g(r,s)$03

then processes it by a coupled $g(r,s)$04D and $g(r,s)$05D UNet. The $g(r,s)$06D propagation assists the $g(r,s)$07D propagation part of the cost volume at different layers by adding visual features to the geometric context, and the final disparity is obtained by soft-argmin. The method is reported as ranked $g(r,s)$08nd on KITTI2012 and ETH3D while being significantly faster than the $g(r,s)$09st-ranked method; on KITTI 2012, the listed runtime is $g(r,s)$10 s for $g(r,s)$11px-all $g(r,s)$12, and on ETH3D the listed runtime is $g(r,s)$13 s for Bad $g(r,s)$14 and Bad $g(r,s)$15 (Kwon et al., 2022).

Stereo reconstruction from non-regular sampling sensors shows yet another integration strategy. Two incomplete high-resolution images are first reconstructed independently by Frequency Selective Extrapolation, then matched by SAD, then real measured sub-pixel samples are transferred from one view into the other before a second extrapolation stage. Compared to a state-of-the-art single-view reconstruction algorithm, the reported average gain in PSNR is $g(r,s)$16 dB; more specifically, the proposed stereo FSE yields an average gain of $g(r,s)$17 dB over the 2006 set $g(r,s)$18 images$g(r,s)$19 and $g(r,s)$20 dB over 2005 $g(r,s)$21 images$g(r,s)$22 (Jonscher et al., 2022).

5. Stereocorrelation in full-field experimental mechanics

In experimental solid mechanics, stereocorrelation is not limited to disparity estimation; it functions as a full-field kinematic measurement method. The projection-based coupling of infrared thermography and stereocorrelation-based digital image correlation starts from independently calibrated systems, synchronized visible-light stereo image pairs and IR images, and stereo correlation to obtain $g(r,s)$23D coordinates of material points. A pinhole camera model with projection matrix $g(r,s)$24 then maps each $g(r,s)$25D point $g(r,s)$26 into the IR image through

$g(r,s)$27

so that temperatures $g(r,s)$28 can be sampled at the projected coordinates and attached to the Lagrangian points as $g(r,s)$29 (Tröger et al., 27 Jun 2026).

The projection matrix can be calibrated by Direct Linear Transformation from $g(r,s)$30 correspondences, by constructing a $g(r,s)$31 linear system $g(r,s)$32, taking the right singular vector associated with the smallest singular value, and reshaping it into the $g(r,s)$33 matrix $g(r,s)$34. The factorization $g(r,s)$35, followed by RQ-decomposition of the left $g(r,s)$36 block and $g(r,s)$37, recovers intrinsic and extrinsic parameters. The same framework proposes a global radial basis function ansatz in both space and time to compute in-plane temperature gradients and even temperature rates on curved surfaces (Tröger et al., 27 Jun 2026).

The frequency-based stereocorrelation method supplies a related but distinct full-field architecture. It evaluates distortion from a single regular grid by phase analysis, without any assumption on the lens model, and applies the inverse warp to all subsequent images. Out-of-plane shape and its variation are measured by temporal phase-stepping fringe projection, with object phase

$g(r,s)$38

and the pin-hole assumption is used both for the projector and the camera to recover full $g(r,s)$39D coordinates. In-plane and out-of-plane changes are then combined by

$g(r,s)$40

The reported application examples include an in-plane tensile test on human skin and $g(r,s)$41D deformation of a leather fabric under tension (Molimard et al., 2013).

This literature makes clear that, in mechanics, stereocorrelation-based DIC is often embedded in a larger measurement chain rather than treated as an isolated stereo algorithm. This suggests a broader interpretation of image stereocorrelation as a registration framework for material points, shape, thermal fields, and derived quantities such as in-plane temperature gradients and rates (Tröger et al., 27 Jun 2026).

6. Applications, performance regimes, and limitations

The surveyed applications are heterogeneous. Mobile stereo-matching systems are described as important in automated-driving vehicles and autonomous robots, where limited hardware resources and low power consumption make it difficult to maintain both acceptable processing speed and accuracy; the embedded-GPU ZNCC work is explicitly aimed at that regime (Chang et al., 2022). Low-resolution thermal stereo is developed in the context of a localization and tracking application and is evaluated on a stereo system manufactured from cheap low-resolution thermal cameras (Zoetgnande et al., 2019). Binocular parallax stereo imaging is demonstrated on a uniform-colored plastic robot holding a sandglass, with measured width and height differing by less than $g(r,s)$42 mm from physical values and a depth range of $g(r,s)$43–$g(r,s)$44 m (Zhu et al., 2014). In solid mechanics, the coupling of DIC and IRT is presented as particularly suited for, but not restricted to, curved surfaces and as straightforward to embed in existing experimental protocols (Tröger et al., 27 Jun 2026).

Several misconceptions are addressed indirectly by the technical designs. One is that stereocorrelation is necessarily a purely local, window-based procedure. The optimal-transport formulation shows that stereo vision can be cast as a single convex program per line, while the ICVP architecture shows that cost-volume stereo can be processed by coupled $g(r,s)$45D and $g(r,s)$46D propagation rather than memory-demanding $g(r,s)$47D cost volumes with large $g(r,s)$48D convolutions alone (Galeotti et al., 2022, Kwon et al., 2022). Another is that stereo correlation depends on strong surface texture. Phase congruency is introduced precisely because thermal images lack texture and suffer from brightness drifts, and correlation-based BPSI is reported to perform better when the object's superficial characteristics are not obvious (Zoetgnande et al., 2019, Zhu et al., 2014).

The limitations are equally explicit. Local correlation methods are described as fast but sensitive to noise and repetitive texture, producing streaking, requiring ad-hoc post-filters, and breaking down at depth discontinuities and occlusion boundaries (Galeotti et al., 2022). In the visual-sensitive-information method, very low-texture regions may not allow the dynamic window to grow far enough, occluded areas still require heuristic hole-filling, and parameter tuning is empirical and image-dependent (Wang et al., 2021). In ST, the computational load of the multiscale, multiorientation filters and the need to tune $g(r,s)$49 versus matching reliability remain current limitations (Zoetgnande et al., 2019). In projection-based DIC–IRT coupling, the accuracy of a single-image DLT and the neglect of lens distortion are limitations, and RBF interpolation can suffer boundary-effect artifacts while noise in IR data can dominate temperature-rate estimates (Tröger et al., 27 Jun 2026). In BPSI, computational complexity is $g(r,s)$50, performance is weaker on image margins and corner points, and depth resolution of a few millimeters lags scanning-based interferometric methods (Zhu et al., 2014).

Across these lines of work, image stereocorrelation is not a single algorithmic primitive but a set of related strategies for establishing correspondence under different sensing assumptions. The shared objective is stable mapping between views or states; the principal differences lie in the correlation observable, the physical image formation model, the treatment of occlusions and smoothness, and the computational substrate on which the correspondence problem is solved.