Pith. sign in

REVIEW 4 major objections 4 minor 57 references

Gaussian kernel-based motion measurement

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Tracking Gaussian kernel locations between frames recovers sub-pixel motion with one fixed parameter set.

desk verdict The paper's headline claim isn't operational: no cross-frame kernel correspondence or motion estimator is defined, so the method can't be executed as written. read the letter →

arxiv 2507.13693 v1 pith:EJTKOPYN submitted 2025-07-18 cs.CV

Hongyi Liu , Haifeng Wang This is my paper
classification cs.CV
keywords sub-pixelmotionmeasurementGaussiankernelconsistencysuper-resolutionconstraintstructuralhealthmonitoringvision-baseddisplacementparameteroptimizationparameter-freetracking
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes a vision-based method for measuring sub-pixel structural motion while avoiding the per-sample parameter tuning required by many existing phase-based and matching techniques. The idea is to represent each video frame as a sum of Gaussian kernels; under the assumption that the measured region translates rigidly, the change in every kernel's center position equals the same displacement vector, so motion is recovered by fitting kernels to two consecutive frames and reading off that vector. To keep the fitted intensity surface from drifting between pixels, the method adds a super-resolution constraint that penalizes errors at interpolated locations. Numerical tests on synthetic patterns and general images report errors mostly within 5 percent of ground-truth displacements from 0.001 to 0.9 pixels, and a laboratory translation test reports mean absolute errors at or below 0.004 mm, all with one default kernel count and one set of loss parameters.

What carries the argument

The load-bearing object is the Gaussian-kernel image representation, $I(\mathbf{X})=\sum_{i=1}^N G_i(\mathbf{X})+E(\mathbf{X})$, where each kernel $G_i$ is a Gaussian brightness bump with center brightness $c_i$, center location $\mu_i$, and covariance set by scale $S_i$ and rotation $R_i$. Kernel parameters are fitted to both frames by gradient-based optimization with an alternating splitting-and-pruning scheme; the representation error $E$ appears in the loss. Two constraints carry the argument: motion consistency, Eq. (6), which says that in a locally rigid region all kernels shift by the same vector $\mathbf{C}$ with no rotation or scale change, turning motion estimation into estimation of $\mathbf{C}$; and the super-resolution constraint $L_s$, which samples random points between pixel-grid locations, penalizes intensity errors there beyond a threshold $\beta$, and prevents the fitted surface from overfitting the pixel samples so that the recovered $\mathbf{C}$ stays accurate. The fixed practical defaults are 3000 kernels, loss weight $w_s=0.33$, threshold $\beta=0.001$, and four interpolated points between pixel pairs.

What would settle it

Take a synthetic or real textured patch, rotate it by a known small angle around its center while keeping the translation zero, and run the method, then compare the recovered $\mathbf{C}$ with the known zero translation and the measured rotation with a known reference. If the method returns a substantial nonzero translation or fails to flag the rotation, the rigid-translation assumption is the reason.

Watch

Extended reading notes

Core claim

The central claim is that sub-pixel motion between two frames can be measured by tracking the locations of Gaussian kernels rather than by matching templates or by filtering at hand-chosen scales. A frame is written as the sum of Gaussian kernels with parameters for brightness, center, scale, and orientation; a deformation from one frame to the next is encoded as changes in those parameters. If the selected region behaves as a rigid body, then all kernel center shifts equal a common vector $\mathbf{C}=(d_x,d_y)^T$ and the rotation and scale changes vanish, so finding the kernels in both frames and taking the mean center displacement estimates the motion. The authors add a super-resolution loss that samples interpolated points between pixels, so the fitted intensity surface cannot overfit the discrete pixel grid and pull the estimated center shift off the true displacement. On synthetic Gaussian-kernel arrays the method reaches $10^{-5}$-pixel errors at 16-bit depth and within about 7 percent of the applied motion at 8-bit depth; on general textured images it stays mostly within 5 percent error for motions from 0.001 to 0.9 pixels; and on a printed-target translation experiment it gives mean absolute errors no greater than 0.004 mm.

Load-bearing premise

The method assumes the selected image region moves as a rigid body: every Gaussian kernel shifts by the same two-dimensional vector, with no rotation, scaling, or non-uniform deformation within that region.

Editorial extensions

If this is right

  • A single default configuration (3000 kernels, $w_s=0.33$, $\beta=0.001$, four interpolated points) suffices for the tested patterns, so no per-image search over pyramid layers, target pixels, or filter parameters is needed.
  • The super-resolution constraint cuts absolute percentage error by up to 97.2 percent in the tested cases and keeps errors around the 5 percent level where the unconstrained fit had errors above 10 percent.
  • Accuracy improves with intensity quantization: 16-bit frames support $10^{-5}$-pixel errors while 8-bit frames stay near $10^{-4}$-pixel errors, indicating that higher-dynamic-range cameras can directly buy finer motion resolution.
  • In the laboratory, the method resolves sub-pixel translations of about 0.32 to 0.44 pixels with mean absolute error at or below 0.004 mm, no more than about 15 percent of the ground-truth motion.
  • Because it reads out only kernel-center shifts, the method is formulated for translational motion of a locally rigid region, which covers typical structural vibration measurement scenarios.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editor's inference: the same kernel representation could be extended to estimate rotation and scaling by relaxing Eq. (6) to fit a shared affine change across kernels, but the estimator would then need an identifiability argument that the paper does not provide.
  • Editor's inference: the random super-resolution sampling acts as a stochastic regularizer; a natural test outside the paper is whether accuracy degrades if the sampling is made deterministic, since the current results do not isolate that effect.
  • Editor's inference: because the motion signal lives in kernel centers, the method might naturally produce dense displacement fields over a grid of patches, potentially supporting full-field vibration analysis with the same parameter set.
  • Editor's inference: the 5 percent error band on 0.001-pixel motions sits close to the physical limits discussed for sub-pixel detection, so applying the method to noisy real videos rather than clean synthetic frames is the decisive next test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a Gaussian-kernel-based method for sub-pixel motion measurement. It decomposes each frame into a set of Gaussian kernels by optimizing a reconstruction loss with an added super-resolution constraint, then claims that the motion of a selected region can be obtained by tracking the positional changes of the kernels, relying on a rigid-body motion-consistency assumption. The authors validate the method on synthetic Gaussian kernel arrays, synthetic general frames with known sub-pixel shifts, and a laboratory translation experiment, reporting errors at the 10^-4-pixel level in simulation and below 0.004 mm in the experiment, all with a fixed set of hyperparameters.

Significance. If the method were fully specified and validated, it could be of practical interest for vision-based structural health monitoring, offering an alternative to phase-based and template-matching approaches that require tedious parameter selection. The paper also usefully identifies the overfitting problem in kernel-based image representation and proposes a plausible super-resolution loss to mitigate it. However, as written, the central mechanism for extracting motion is not defined: the optimization objective contains no correspondence term or estimator that converts kernel parameters into a translation vector, and the primary simulation is a self-referential test in which the data are generated by the same Gaussian model the method fits. The external validations do not compensate for these gaps because the GFSM motion synthesis shares a linear-interpolation model with the proposed super-resolution loss, and the hyperparameters are selected using the same test samples. The paper therefore does not substantiate its main claim of a parameter-free, accurate tracking method.

major comments (4)
  1. [Section 2.2, Eq. (7)] The motion-consistency assumption in Eq. (6) is never imposed in the optimization. The loss L in Eq. (7) contains only per-frame reconstruction errors and the super-resolution term L_s; there is no term that links the i-th kernel in the current frame to the i-th kernel in the next frame, no penalty on δR_i or δS_i, and no term pulling δµ_i toward a common translation C. Because Gaussian kernels are unordered and the splitting/pruning procedure is randomly initialized, the index i in Eqs. (5) and (6) is arbitrary. Moreover, the paper does not specify how the translation C is estimated from the optimized kernel parameters (mean of δµ_i? median? robust fit?). Consequently, the reported MAEs in Tables 2 and 3 do not follow from the stated equations or algorithm.
  2. [Section 3.1, Table 2] The GKA validation is circular. Frames are generated as exact sums of Gaussian kernels with known centers, scales, and orientations, and the method fits the same Gaussian model to those frames; the ground-truth motion is applied directly to the kernel centers. This experiment tests the optimizer's ability to recover parameters of its own generative model, not its ability to measure motion in arbitrary image content. The extremely small MAEs in Table 2 are therefore not evidence for the abstract's general claim of high accuracy.
  3. [Section 3.2 and Eq. (8)] The GFSM validation is favorably biased by the shared use of linear interpolation. Synthetic motions in Section 3.2 are generated by shifting intensity according to the linear relation ΔI = I·d, and the super-resolution loss L_s in Eq. (8) is defined as the error between the rendered frame and the linearly interpolated target frame. The loss thus rewards exactly the interpolation model used to create the test data, which can inflate the measured accuracy and does not support the method's generality beyond this specific motion-synthesis model.
  4. [Section 3.2.1 and abstract] The claim of operation 'without customized parameter setup' is not established. The kernel number N=3000 is selected from a grid search on the GFSM images (Fig. 8), and the values w_s=0.33, β=0.001, and four interpolated points are fixed without a principled derivation or sensitivity analysis. Because these choices are evaluated on the same type of samples used in the main tests, the results do not demonstrate that the method is parameter-free; they only show that one set of hand-picked parameters was held fixed for the reported experiments.
minor comments (4)
  1. [Section 4.2, Table 3] The MAE entries in Table 3 appear to repeat and shift across rows (for example, 0.0015/0.0000 appears as the second entry of GKA region 1 and as the first entry of GKA region 2, and the pattern continues downward). If this is not a copy-and-paste error, the table needs a clearer description; if it is an error, it should be corrected because it undermines confidence in the experimental reporting.
  2. [Section 3.2.1] The text contains a typo: 'super-solution pixels' should read 'super-resolution pixels'.
  3. [Section 3.2] The description of how the test images are down-sampled and how sub-regions are extracted lacks details (downsampling factor, interpolation method, region selection criterion), which makes the numerical experiments difficult to reproduce.
  4. [Section 4.1] The experimental validation reports only three ground-truth motion levels and does not provide uncertainty estimates for the laser sensor or the camera's pixel-to-millimeter calibration, so the reported MAEs should be interpreted cautiously.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity: the motion estimate is not fed into the loss, and the lab experiment is externally grounded; the only mild self-consistency is the shared linear-interpolation model in the GFSM validation and the super-resolution constraint.

  1. other [Section 2.2.2 (Eq. 8) and Section 3.2 (GFSM synthesis)]
    "the error E(X ls,ms) is calculated as the difference between rendered frame and the linearly interpolated target frame (Fig. 3) ... we adopted a linear interpolation technique described in Miao et al. [52]. This method simulates sub-pixel motion by shifting a proportion of the pixel intensity to the adjacent pixel, following the relationship ∆I = I · d."

    The super-resolution regularizer Ls is defined against a linearly interpolated target frame, while the GFSM ground-truth frames are generated by exactly the same linear-interpolation motion model. Thus the numerical validation partly rewards the method for conforming to its own modeling assumption rather than testing motion tracking against an independent deformation model. This does not force the motion estimate itself, because the loss never uses the ground-truth shift, but it makes the GFSM accuracy numbers partially self-consistent and therefore less compelling as independent evidence.

full rationale

The core derivation is not circular. Equation (4) defines frames as sums of Gaussian kernels; Equations (5)-(6) define parameter variations and the rigid-body assumption; Equations (7)-(8) define a reconstruction loss that contains no ground-truth motion term. The reported motion C is a derived quantity from fitted kernel parameters, not a fitted target, so the central claim is not equivalent to the inputs by construction. The GKA test is an in-family synthetic test but is still an inverse problem: the optimizer is not given the known shifts and must recover them from rendered frames, so it is not a logical reduction. The skeptical concern that the paper never specifies how kernel correspondences are established or how C is estimated from δµ_i is a real operational gap, but that is a completeness/correctness issue rather than circularity. There are no load-bearing author self-citations; the cited Kernel-splatting and linear-interpolation works are external. The experimental validation against a laser displacement sensor provides independent, out-of-sample support. The only notable circularity-adjacent element is the shared linear-interpolation model between the super-resolution loss and the GFSM synthesis, which biases the numerical validation but does not force the central claim. Overall circularity is therefore mild.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central method relies on the representational power of Gaussian sums and a rigid-body assumption, plus several hand-chosen hyperparameters (kernel number, loss weight, threshold, interpolation density). No code or data is provided, and the motion-consistency mechanism is not specified.

free parameters (5)
  • Kernel number N = 3000 (2000 for GKA)
    Selected from the kernel-number sweep in Section 3.2.1 to minimize error; applied as a fixed default to all samples.
  • Super-resolution loss weight ws = 0.33
    Set by hand; no sensitivity analysis or theoretical motivation is given in Section 2.2.2.
  • Super-resolution threshold beta = 0.001
    Set by hand so that only errors above this level contribute to the loss (Eq. 8).
  • Interpolated points per pixel pair = 4
    Chosen for the super-resolution grid; no study of its effect is provided.
  • Super-resolution sampling fraction = 5%
    Chosen for computational efficiency; used in Eq. (8) but not analyzed.
assumptions (4)
  • domain assumption A video frame can be represented as a sum of a finite number of 2D Gaussian kernels with acceptable error (Eq. 4).
    The method's accuracy depends on how well Gaussian sums approximate arbitrary image content; no error bound is provided.
  • domain assumption The selected region is a rigid body undergoing uniform translation, so all kernel centers shift by the same C and scales/orientations are unchanged (Eq. 6).
    This is stated as fitting practical structural conditions but is not verified on deformable or rotating targets.
  • standard math The alternative-splitting-and-pruning kernel optimization from Kerbl et al. [55] converges to a good representation of both frames.
    The method relies on existing 3D Gaussian splatting optimization; no convergence or sensitivity analysis is given.
  • domain assumption Linear interpolation correctly synthesizes sub-pixel motion in GFSM tests (Section 3.2).
    The synthetic ground truth is generated by the same interpolation model used to define super-resolution error, which could bias results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Gaussian kernel-based motion measurement." pith.science (2026). https://pith.science/paper/EJTKOPYN

@misc{pith2026250713693,
  author       = {Pith},
  title        = {Pith review of: Gaussian kernel-based motion measurement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EJTKOPYN}},
  note         = {Machine review of arXiv:2507.13693}
}
read the original abstract

The growing demand for structural health monitoring has driven increasing interest in high-precision motion measurement, as structural information derived from extracted motions can effectively reflect the current condition of the structure. Among various motion measurement techniques, vision-based methods stand out due to their low cost, easy installation, and large-scale measurement. However, when it comes to sub-pixel-level motion measurement, current vision-based methods either lack sufficient accuracy or require extensive manual parameter tuning (e.g., pyramid layers, target pixels, and filter parameters) to reach good precision. To address this issue, we developed a novel Gaussian kernel-based motion measurement method, which can extract the motion between different frames via tracking the location of Gaussian kernels. The motion consistency, which fits practical structural conditions, and a super-resolution constraint, are introduced to increase accuracy and robustness of our method. Numerical and experimental validations show that it can consistently reach high accuracy without customized parameter setup for different test samples.

Figures

Figures reproduced from arXiv: 2507.13693 by the authors.

Figure 1
Figure 1. How to stack various Gaussian kernels to represent an image [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An example of overfitting on one intensity line of a frame [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Sample region for super-resolution locations (Only insert 1 interpolated point between each pair of pixels [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Gaussian kernel array 5 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Motion MAE for GKA 6 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Test sample images with sub-pixel motions [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Diagram of motion synthesis method 3.2.1 Kernel number The number of kernels used to represent the entire frame has a notable impact on motion measurement accuracy. Using too few kernels may lead to an under-representation of image content and the lack of fine details,…
Figure 8
Figure 8. Figure 8: Motion MAE for GFSM with 0.01-pixel motion on different kernel numbers [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Motion MAE for GFSM with/without Ls A significant improvement in motion measurement accuracy is observed in most test cases when incorporating the super-resolution constraint Ls with an improvement on the percentage error of from 5.8% up to 97.2%. These results highlig…
Figure 2
Figure 2. Figure 2: As a result, while the super-resolution constraint reduces the over-fitting issue, it does not significantly increase [PITH_FULL_IMAGE:figures/full_fig_p010_2.png]
Figure 10
Figure 10. Figure 10: Motion MAE for GFSM with different motion values [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Experiment setup for the micrometer-driven translation stage [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Selected regions from the frame captured by DJI Osmo Action 5 Pro [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 56 canonical work pages

  1. [1]

    Review on vibration-based structural health monitoring techniques and technical codes

    Yang Yang, Yao Zhang, and Xiaokun Tan. Review on vibration-based structural health monitoring techniques and technical codes. Symmetry, 13(11):1998, October 2021

    work page 1998

  2. [2]

    Research on traditional and deep learning strategies based on optical flow estimation - a review

    Yifan Wang, Wu Wang, Yang Li, Jinshi Guo, Yu Xu, Jiaqi Ma, Yu Ling, Yanan Fu, and Yaodong Jia. Research on traditional and deep learning strategies based on optical flow estimation - a review. Journal of King Saud University - Computer and Information Sciences, 36(4):102029, April 2024

    work page 2024

  3. [3]

    Fleet and Y

    D. Fleet and Y . Weiss. Optical flow estimation. In Nikos Paragios, Yunmei Chen, and Olivier Faugeras, editors, Handbook of Mathematical Models in Computer Vision, pages 237–257. Springer-Verlag, New York, 2006

    work page 2006

  4. [4]

    Picture processing by computer

    Azriel Rosenfeld. Picture processing by computer. ACM Computing Surveys, 1(3):147–176, September 1969

    work page 1969

  5. [5]

    Weng, T.S

    J. Weng, T.S. Huang, and N. Ahuja. Motion from images: image matching, parameter estimation and intrinsic stability. In [1989] Proceedings. Workshop on Visual Motion , pages 359–366, Irvine, CA, USA, 1989. IEEE Comput. Soc. Press

    work page 1989

  6. [6]

    Yan Xu and James M. W. Brownjohn. Review of machine-vision based methodologies for displacement mea- surement in civil structures. Journal of Civil Structural Health Monitoring, 8(1):91–110, January 2018

    work page 2018

  7. [7]

    A survey of feature matching methods

    Qian Huang, Xiaotong Guo, Yiming Wang, Huashan Sun, and Lijie Yang. A survey of feature matching methods. IET Image Processing, 18(6):1385–1410, May 2024

    work page 2024

  8. [8]

    Motion compensated inter-frame coding for video conferencing

    Toshio Koga. Motion compensated inter-frame coding for video conferencing. 1981. 13

    work page 1981

Show all 57 references
  1. [9]

    Template matching advances and applications in image analysis, 2016

    Nazanin Sadat Hashemi, Roya Babaie Aghdam, Atieh Sadat Bayat Ghiasi, and Parastoo Fatemi. Template matching advances and applications in image analysis, 2016. Version Number: 1

    2016

  2. [10]

    Khawase, Shailesh D

    Sonam T. Khawase, Shailesh D. Kamble, Nileshsingh V . Thakur, and Akshay S. Patharkar. An overview of block matching algorithms for motion vector estimation. pages 217–222, June 2017

    2017

  3. [11]

    Improved block matching motion estimation technique using modified particle swarm optimiza- tion in video coding

    Deepak Singh. Improved block matching motion estimation technique using modified particle swarm optimiza- tion in video coding. In 2021 Emerging Trends in Industry 4.0 (ETI 4.0), pages 1–6, Raigarh, India, May 2021. IEEE

    2021

  4. [12]

    A novel four-step search algorithm for fast block motion estimation

    Lai-Man Po and Wing-Chung Ma. A novel four-step search algorithm for fast block motion estimation. IEEE Transactions on Circuits and Systems for Video Technology, 6(3):313–317, June 1996

    1996

  5. [13]

    A new diamond search algorithm for fast block-matching motion estimation

    Shan Zhu and Kai-Kuang Ma. A new diamond search algorithm for fast block-matching motion estimation. IEEE Transactions on Image Processing, 9(2):287–290, February 2000

    2000

  6. [14]

    A novel hexagon-based search algorithm for fast block motion estimation

    Ce Zhu, Xiao Lin, Lap-Pui Chau, Keng-Pang Lim, Hock-Ann Ang, and Choo-Yin Ong. A novel hexagon-based search algorithm for fast block motion estimation. In 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221) , volume 3...

    2001

  7. [15]

    Necati Catbas

    Tung Khuc and F. Necati Catbas. Computer vision-based displacement and vibration monitoring without using physical target on structures. Structure and Infrastructure Engineering, 13(4):505–516, April 2017

    2017

  8. [16]

    A non-target structural displacement measurement method using advanced feature matching strategy

    Chuan-Zhi Dong and F Necati Catbas. A non-target structural displacement measurement method using advanced feature matching strategy. Advances in Structural Engineering, 22(16):3461–3472, December 2019

    2019

  9. [17]

    D.G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the Seventh IEEE Interna- tional Conference on Computer Vision, pages 1150–1157 vol.2, Kerkyra, Greece, 1999. IEEE

    1999

  10. [18]

    Sukthankar

    Yan Ke and R. Sukthankar. PCA-SIFT: a more distinctive representation for local image descriptors. InProceed- ings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., volume 2, pages 506–513, Washington, DC, USA, 2004. IEEE

    2004

  11. [19]

    SURF: Speeded up robust features

    Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: Speeded up robust features. In Aleš Leonardis, Horst Bischof, and Axel Pinz, editors, Computer Vision – ECCV 2006, volume 3951, pages 404–417. Springer Berlin Heidelberg, Berlin, Heidelberg, 2006. Series Title: Lecture Not...

    2006

  12. [20]

    BRIEF: Binary robust independent elementary features

    Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. BRIEF: Binary robust independent elementary features. In David Hutchison, Takeo Kanade, Josef Kittler, Jon M. Kleinberg, Friedemann Mattern, John C. Mitchell, Moni Naor, Oscar Nierstrasz, C. Pandu Rangan, Be...

    2010

  13. [21]

    E. Tola, V . Lepetit, and P. Fua. DAISY: An efficient dense descriptor applied to wide-baseline stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(5):815–830, May 2010

    2010

  14. [22]

    ORB: An efficient alternative to SIFT or SURF

    Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In 2011 International Conference on Computer Vision, pages 2564–2571, Barcelona, Spain, November

    2011

  15. [23]

    Siegwart

    Stefan Leutenegger, Margarita Chli, and Roland Y . Siegwart. BRISK: Binary robust invariant scalable keypoints. In 2011 International Conference on Computer Vision , pages 2548–2555, Barcelona, Spain, November 2011. IEEE

    2011

  16. [24]

    Pablo Fernández Alcantarilla, Adrien Bartoli, and Andrew J. Davison. KAZE features. In David Hutchison, Takeo Kanade, Josef Kittler, Jon M. Kleinberg, Friedemann Mattern, John C. Mitchell, Moni Naor, Oscar Nier- strasz, C. Pandu Rangan, Bernhard Steffen, Madhu Sudan, Demetri T...

    2012

  17. [25]

    Alahi, R

    A. Alahi, R. Ortiz, and P. Vandergheynst. FREAK: Fast retina keypoint. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 510–517, Providence, RI, June 2012. IEEE

    2012

  18. [26]

    LIFT: Learned invariant feature transform,

    Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned invariant feature transform,

  19. [27]

    L2-Net: Deep learning of discriminative patch descriptor in Euclidean space

    Yurun Tian, Bin Fan, and Fuchao Wu. L2-Net: Deep learning of discriminative patch descriptor in Euclidean space. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 6128–6136, Honolulu, HI, July 2017. IEEE. 14

    2017

  20. [28]

    Working hard to know your neighbor’s margins: Local descriptor learning loss, 2017

    Anastasiya Mishchuk, Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Working hard to know your neighbor’s margins: Local descriptor learning loss, 2017. Version Number: 4

    2017

  21. [29]

    GeoDesc: Learning local descriptors by integrating geometry constraints

    Zixin Luo, Tianwei Shen, Lei Zhou, Siyu Zhu, Runze Zhang, Yao Yao, Tian Fang, and Long Quan. GeoDesc: Learning local descriptors by integrating geometry constraints. 2018. Publisher: arXiv Version Number: 2

    2018

  22. [30]

    Structural dynamic displacement vision system using digital image processing

    Hyoung-Suk Choi, Jin-Hwan Cheung, Sang-Hyo Kim, and Jin-Hee Ahn. Structural dynamic displacement vision system using digital image processing. NDT & E International, 44(7):597–608, November 2011

    2011

  23. [31]

    Development of a vision based deflection measurement system and its accuracy assessment

    Jerzy Sładek, Ksenia Ostrowska, Piotr Kohut, Krzysztof Holak, Adam G ˛ aska, and Tadeusz Uhl. Development of a vision based deflection measurement system and its accuracy assessment. Measurement, 46(3):1237–1249, April 2013

    2013

  24. [32]

    EpicFlow: Edge-preserving inter- polation of correspondences for optical flow

    Jerome Revaud, Philippe Weinzaepfel, Zaid Harchaoui, and Cordelia Schmid. EpicFlow: Edge-preserving inter- polation of correspondences for optical flow. 2015

    2015

  25. [33]

    Reddy and B.N

    B.S. Reddy and B.N. Chatterji. An FFT-based technique for translation, rotation, and scale-invariant image registration. IEEE Transactions on Image Processing, 5(8):1266–1271, August 1996

    1996

  26. [34]

    Novel template matching method with sub-pixel accuracy based on correlation and Fourier-Mellin transform

    Guangjun Zhang, Ming Lei, and Xulin Liu. Novel template matching method with sub-pixel accuracy based on correlation and Fourier-Mellin transform. Optical Engineering, 48(5):057001, May 2009

    2009

  27. [35]

    Robust fine registration of multisensor remote sensing images based on enhanced subpixel phase correlation

    Zhen Ye, Jian Kang, Jing Yao, Wenping Song, Sicong Liu, Xin Luo, Yusheng Xu, and Xiaohua Tong. Robust fine registration of multisensor remote sensing images based on enhanced subpixel phase correlation. Sensors, 20(15):4338, August 2020

    2020

  28. [36]

    Image-to-image subpixel registration based on template matching of road network extracted by deep learning

    Shuhei Hikosaka and Hideyuki Tonooka. Image-to-image subpixel registration based on template matching of road network extracted by deep learning. Remote Sensing, 14(21):5360, October 2022

    2022

  29. [37]

    Realistic limits for subpixel movement detection

    David Mas, Jorge Perez, Belen Ferrer, and Julian Espinosa. Realistic limits for subpixel movement detection. Applied Optics, 55(19):4974, July 2016

    2016

  30. [38]

    An iterative image registration technique with an application to stereo vision

    Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. IJCAI’81: 7th international joint conference on Artificial intelligence, pages 674–679, August 1981

    1981

  31. [39]

    Horn and Brian G

    Berthold K.P. Horn and Brian G. Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185–203, August 1981

    1981

  32. [40]

    Chen, Neal Wadhwa, Young-Jin Cha, Frédo Durand, William T

    Justin G. Chen, Neal Wadhwa, Young-Jin Cha, Frédo Durand, William T. Freeman, and Oral Buyukozturk. Modal identification of simple structures with high-speed video using motion magnification. Journal of Sound and Vibration, 345:58–71, June 2015

    2015

  33. [41]

    Deqing Sun, Stefan Roth, and Michael J. Black. Secrets of optical flow estimation and their principles. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2432–2439, San Francisco, CA, USA, June 2010. IEEE

    2010

  34. [42]

    Optic flow in harmony

    Henning Zimmer, Andrés Bruhn, and Joachim Weickert. Optic flow in harmony. International Journal of Computer Vision, 93(3):368–388, July 2011

    2011

  35. [43]

    Optical flow modeling and computation: A survey

    Denis Fortun, Patrick Bouthemy, and Charles Kervrann. Optical flow modeling and computation: A survey. Computer Vision and Image Understanding, 134:1–21, May 2015

    2015

  36. [44]

    Regularization strategies for discontinuity-preserving optical flow methods

    Nelson Monzon, Agustin Salgado, and Javier Sanchez. Regularization strategies for discontinuity-preserving optical flow methods. IEEE Transactions on Image Processing, 25(4):1580–1591, April 2016

    2016

  37. [45]

    Veltkamp, Baoxin Li, and Junsong Yuan

    Zhigang Tu, Wei Xie, Dejun Zhang, Ronald Poppe, Remco C. Veltkamp, Baoxin Li, and Junsong Yuan. A survey of variational and CNN-based optical flow techniques. Signal Processing: Image Communication, 72:9– 24, March 2019

    2019

  38. [46]

    Fleet and Allan D

    David J. Fleet and Allan D. Jepson. Computation of component image velocity from local phase information. International Journal of Computer Vision, 5(1):77–104, August 1990

    1990

  39. [47]

    Gautama and M.A

    T. Gautama and M.A. Van Hulle. A phase-based approach to the estimation of the optical flow field using spatial filtering. IEEE Transactions on Neural Networks, 13(5):1127–1136, September 2002

    2002

  40. [48]

    Diamond, P.S

    D.H. Diamond, P.S. Heyns, and A.J. Oberholster. Accuracy evaluation of sub-pixel structural vibration measure- ments through optical flow analysis of a video sequence. Measurement, 95:166–172, January 2017

    2017

  41. [49]

    Liu, M.Z

    G. Liu, M.Z. Li, Z. Mao, and Q.S. Yang. Structural motion estimation via Hilbert transform enhanced phase- based video processing. Mechanical Systems and Signal Processing, 166:108418, March 2022

    2022

  42. [50]

    Phase-based displacement measurement on a straight edge using an optimal complex Gabor filter

    Yinan Miao, Jun Young Jeon, Yeseul Kong, and Gyuhae Park. Phase-based displacement measurement on a straight edge using an optimal complex Gabor filter. Mechanical Systems and Signal Processing , 164:108224, February 2022. 15

    2022

  43. [51]

    Phase-based vibration imaging for structural dynamics applications: marker-free full-field displacement measurements with confidence mea- sures

    Yinan Miao, Yeseul Kong, Hyeonwoo Nam, Seunghwan Lee, and Gyuhae Park. Phase-based vibration imaging for structural dynamics applications: marker-free full-field displacement measurements with confidence mea- sures. Mechanical Systems and Signal Processing, 198:110418, September 2023

    2023

  44. [52]

    A novel marker for robust and accurate phase-based 2d motion estimation from noisy image data

    Yinan Miao, Yeseul Kong, Jun Young Jeon, Hyeonwoo Nam, and Gyuhae Park. A novel marker for robust and accurate phase-based 2d motion estimation from noisy image data. Mechanical Systems and Signal Processing, 187:109931, March 2023

    2023

  45. [53]

    Vibration measure- ment from an adaptive phase-based motion estimation using parameter optimised log-Gabor filter.Measurement, 224:113812, January 2024

    Wendi Zhang, Xiaojian Wang, Hongguang Li, Jiwen Zhou, Zhong Luo, and Guang Meng. Vibration measure- ment from an adaptive phase-based motion estimation using parameter optimised log-Gabor filter.Measurement, 224:113812, January 2024

    2024

  46. [54]

    Alternatives to voxels for image representation in iterative reconstruction algorithms

    R M Lewitt. Alternatives to voxels for image representation in iterative reconstruction algorithms. Physics in Medicine and Biology, 37(3):705–716, March 1992

    1992

  47. [55]

    3D Gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1–14, August 2023

    2023

  48. [56]

    Gabor splatting for high-quality gigapixel image representa- tions

    Skylar Wurster, Ran Zhang, and Changxi Zheng. Gabor splatting for high-quality gigapixel image representa- tions. In ACM SIGGRAPH 2024 Posters, pages 1–2, Denver CO USA, July 2024. ACM

    2024

  49. [57]

    4D Gaussian splatting for real-time dynamic scene rendering, December 2023

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D Gaussian splatting for real-time dynamic scene rendering, December 2023. arXiv:2310.08528 [cs]. 16

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.