Back

In Defense of the Eight-Point Algorithm

R. Hartley

article1997TPAMI1,926 citations

Demonstrates that simple coordinate normalization overcomes the numerical instability of the classic eight-point algorithm, enabling fast and direct fundamental matrix estimation that rivals complex iterative methods in accuracy.

Listen

In computer vision, computing geometric relationships between uncalibrated camera views is essential for three-dimensional reconstruction, stereo matching, and image rectification. The classic linear eight-point method provides a simple, direct mathematical solution to this problem. However, prevailing industry and academic consensus long dismissed it as practically useless due to severe instability in the presence of noise, prompting widespread adoption of much more complex iterative algorithms.

The article sets out to demonstrate that the failure of the eight-point method stems from poor numerical conditioning rather than fundamental theoretical flaws, and that a simple coordinate normalization step restores its accuracy to a level comparable with sophisticated iterative techniques.

To evaluate this proposition, the author tested the modified approach using thousands of randomized trials across five diverse real-world image pairs ranging from accurate calibration rigs to noisy outdoor scenes. The evaluated method introduces an elementary data preprocessing step: translating matched image points to center their coordinates at the origin and scaling them so their average distance from the center equals the square root of two. The fundamental geometric matrix is then calculated linearly, constrained to proper rank using singular value decomposition, and transformed back into the original coordinate system.

The experimental findings show that input normalization improves the numerical condition number of the underlying linear system by a factor of roughly one hundred million. This dramatic reduction in sensitivity reduces calculation errors from as high as ten pixels down to sub-pixel accuracy, matching the precision of the most complex, computationally intensive iterative methods. Furthermore, simple isotropic scaling performs just as effectively as more intricate non-isotropic methods, while running approximately twenty times faster than iterative optimization schemes.

These results demonstrate that engineering teams can achieve state-of-the-art accuracy in multi-view geometry without paying the steep computational and code-maintenance costs associated with complex optimization routines. The findings also highlight that standard iterative algorithms are prone to local minima and often fail unless initialized with properly normalized estimates, making pre-normalization a critical design factor across linear and non-linear implementations alike.

Organizations developing spatial computing, robotics, or computer vision pipelines should immediately adopt coordinate normalization before applying any linear estimation techniques. When iterative solvers are strictly required for maximum precision, developers should use the normalized linear method as the initialization step to prevent convergence failures.

The conclusions are supported with high statistical confidence across diverse real-image benchmarks and noise levels. Readers should note, however, that the approach assumes prior removal of severe outliers and is evaluated primarily on uncalibrated camera setups, though the underlying mathematical normalization principles apply broadly to related computer vision tasks.

General Electric Companyadaptive algorithmsoptimization problemseight-point algorithmepipolar geometrypoint correspondences
Hartley (1997).pdf

No sufficiently relevant recommendations were found.

Cover for In Defense of the Eight-Point Algorithm

Abstract

The fundamental matrix is a basic tool in the analysis of scenes taken with two uncalibrated cameras, and the 8-point algorithm is a frequently cited method for computing the fundamental matrix from a set of 8 or more point matches. It has the advantage of simplicity of implementation. The prevailing view is, however, that it is extremely susceptible to noise and hence virtually useless for most purposes. This paper challenges that view, by showing that by preceding the algorithm with a very simple normalization (translation and scaling) of the coordinates of the matched points, results are obtained comparable with the best iterative algorithms. This improved performance is justified by theory and verified by extensive experiments on real images.

Topics

adaptive algorithmsoptimization problems

Keywords

eight-point algorithmepipolar geometrypoint correspondencesprojective reconstructionsingular value decompositionisotropic scalingsingularity constraintleast-squares solutions

Table of Contents

  • 1 Introduction
  • 2 Outline of the 8-point Algorithm
  • 3 Transformation of the Input
  • 4 Condition of the System of Equations
  • 5 Normalizing transformations
  • 5.1 Isotropic Scaling
  • 5.2 Non-isotropic Scaling
  • 6 Scaling in Stage 2
  • 7 Experimental Evaluation
  • 7.1 Other algorithms.
  • 7.2 The Images.
  • 7.3 Graphical Presentation of the Results.
  • 8 Conclusions
  • 9 Acknowledgements
  • References

Knowls

  1. Knowl 1 — The Normalized 8-Point Algorithm for Fundamental Matrix Estimation

    algorithm

    The normalized 8-point algorithm computes a rank-2 fundamental matrix FF satisfying ui′TFui≈0\mathbf{u}'_i{}^T F \mathbf{u}_i \approx 0 from a set of N≥8N \ge 8 matched point correspondences ui↔ui′\mathbf{u}_i \leftrightarrow \mathbf{u}'_i in two uncalibrated images.

    Input: A set of N≥8N \ge 8 point correspondences {ui↔ui′}i=1N\{\mathbf{u}_i \leftrightarrow \mathbf{u}'_i\}_{i=1}^N, where ui=(ui,vi,1)T\mathbf{u}_i = (u_i, v_i, 1)^T and ui′=(ui′,vi′,1)T\mathbf{u}'_i = (u'_i, v'_i, 1)^T.
    Output: A 3×33 \times 3 fundamental matrix FF of rank 2 satisfying det⁡F=0\det F = 0.
    1. Coordinate Normalization:
       Compute centroid (uˉ,vˉ)(\bar{u}, \bar{v}) of points {ui}i=1N\{\mathbf{u}_i\}_{i=1}^N:
         uˉ=1N∑i=1Nui\bar{u} = \frac{1}{N} \sum_{i=1}^N u_i, vˉ=1N∑i=1Nvi\bar{v} = \frac{1}{N} \sum_{i=1}^N v_i
       Compute the mean Euclidean distance to the centroid:
         davg=1N∑i=1N(ui−uˉ)2+(vi−vˉ)2d_{\text{avg}} = \frac{1}{N} \sum_{i=1}^N \sqrt{(u_i - \bar{u})^2 + (v_i - \bar{v})^2}
       Construct the normalization transformation matrix TT with scaling factor s=2davgs = \frac{\sqrt{2}}{d_{\text{avg}}}:
         T=[s0−suˉ0s−svˉ001]T = \begin{bmatrix} s & 0 & -s\bar{u} \\ 0 & s & -s\bar{v} \\ 0 & 0 & 1 \end{bmatrix}
       Compute transformation matrix T′T' similarly for the second image points {ui′}i=1N\{\mathbf{u}'_i\}_{i=1}^N.
       Normalize points: u^i=Tui\hat{\mathbf{u}}_i = T \mathbf{u}_i and u^i′=T′ui′\hat{\mathbf{u}}'_i = T' \mathbf{u}'_i for each i=1,…,Ni = 1, \dots, N.
    2. Linear Least-Squares Solution (Stage 1):
       Form the N×9N \times 9 equation matrix A^\hat{A}, where each row ii is:
         (u^iu^i′,  u^iv^i′,  u^i,  v^iu^i′,  v^iv^i′,  v^i,  u^i′,  v^i′,  1)(\hat{u}_i \hat{u}'_i, \; \hat{u}_i \hat{v}'_i, \; \hat{u}_i, \; \hat{v}_i \hat{u}'_i, \; \hat{v}_i \hat{v}'_i, \; \hat{v}_i, \; \hat{u}'_i, \; \hat{v}'_i, \; 1)
       Compute the Singular Value Decomposition (SVD) of A^=UADAVAT\hat{A} = U_A D_A V_A^T.
       Extract the 9-vector f^\hat{\mathbf{f}} corresponding to the column of VAV_A associated with the smallest singular value.
       Reshape f^=(f^1,…,f^9)T\hat{\mathbf{f}} = (\hat{f}_1, \dots, \hat{f}_9)^T into the 3×33 \times 3 matrix F^\hat{F} row by row.
    3. Enforce Singularity Constraint (Stage 2):
       Compute the SVD of F^=Udiag⁡(r,s,t)VT\hat{F} = U \operatorname{diag}(r, s, t) V^T, where r≥s≥tr \ge s \ge t.
       Replace F^\hat{F} with the closest rank-2 matrix under the Frobenius norm:
         F^′=Udiag⁡(r,s,0)VT\hat{F}' = U \operatorname{diag}(r, s, 0) V^T
    4. Denormalization:
       F=T′TF^′TF = T'^T \hat{F}' T
       return FF

  2. Knowl 2 — Numerical Ill-Conditioning of the Unnormalized Epipolar Linear System

    theoretical result

    In the standard 8-point algorithm without coordinate normalization, typical image coordinates of magnitude (100,100,1)T(100, 100, 1)^T cause severe ill-conditioning in the normal equation matrix ATAA^T A.

    Because each row of the design matrix AA contains products of homogeneous image coordinates (uu′,uv′,u,vu′,vv′,v,u′,v′,1)(u u', u v', u, v u', v v', v, u', v', 1), the diagonal entries of ATAA^T A are proportional to (108,108,104,108,108,104,104,104,1)(10^8, 10^8, 10^4, 10^8, 10^8, 10^4, 10^4, 10^4, 1), spanning eight orders of magnitude.

    By the Interlacing Property for the eigenvalues of symmetric matrices, the eigenvalues of the trailing 2×22 \times 2 principal submatrix X2X_2 of ATAA^T A bound the eigenvalues of the full 9×99 \times 9 matrix X9=ATAX_9 = A^T A. Specifically: λ8(X9)≤λ1(X2)≤trace⁡(X2)=104+1\lambda_8(X_9) \le \lambda_1(X_2) \le \operatorname{trace}(X_2) = 10^4 + 1 Because the largest eigenvalue λ1(X9)\lambda_1(X_9) is bounded below by the largest diagonal entry (10810^8), the condition ratio κ=λ1(X9)/λ8(X9)\kappa = \lambda_1(X_9) / \lambda_8(X_9) satisfies: κ≥108104+1≈104\kappa \ge \frac{10^8}{10^4 + 1} \approx 10^4 In practice, λ8(X9)\lambda_8(X_9) is far smaller, and the condition number routinely reaches 101110^{11} to 101310^{13}. Under this ill-conditioning, perturbations to ATAA^T A of the order of magnitude of d8d_8 cause large rotations of the least eigenvector subspace, rendering the unnormalized linear solution highly unstable under image noise. Isotropic coordinate normalization reduces the condition number by approximately 10810^8.

  3. Knowl 3 — Relative Perturbation Vulnerability during Singularity Enforcement

    theoretical result

    Enforcing the rank-2 constraint det⁡F=0\det F = 0 via truncated Singular Value Decomposition (SVD) without coordinate normalization disproportionately distorts the geometrically most critical parameters of the fundamental matrix.

    For unnormalized coordinates of magnitude ≈100\approx 100, entries of the resulting fundamental matrix FF span several orders of magnitude: F∼[10−410−410−210−410−410−210−210−21]F \sim \begin{bmatrix} 10^{-4} & 10^{-4} & 10^{-2} \\ 10^{-4} & 10^{-4} & 10^{-2} \\ 10^{-2} & 10^{-2} & 1 \end{bmatrix} When evaluating the epipolar line l′=Fu\mathbf{l}' = F \mathbf{u} for a point u=(u,v,1)T\mathbf{u} = (u, v, 1)^T, the largest coordinate components (u,v≈100u, v \approx 100) multiply the smallest matrix entries (the 10−410^{-4} top-left block). Hence, the smallest numerical entries of FF dominate epipolar line placement.

    Replacing FF with the closest singular matrix F′F' under the Frobenius norm perturbs all matrix entries by roughly equal absolute amounts. Consequently, an absolute perturbation ϵ\epsilon represents an insignificant relative change for entries of magnitude 11, but an immense relative distortion (up to 104ϵ10^4 \epsilon) for the 10−410^{-4} entries. Pre-normalizing image coordinates ensures that all entries of FF have approximately equal magnitude, preventing disproportionate relative distortion during rank-2 enforcement.

  4. Knowl 4 — Non-Equivalence of Least-Squares Epipolar Solutions under Coordinate Transformations

    theoretical result

    Linear coordinate transformations applied to image points alter the algebraic least-squares solution for the fundamental matrix because algebraic distance minimization is not invariant under non-orthogonal transformations.

    Let image coordinates be transformed as u^=Tu\hat{\mathbf{u}} = T \mathbf{u} and u^′=T′u′\hat{\mathbf{u}}' = T' \mathbf{u}'. The epipolar condition u′TFu=0\mathbf{u}'^T F \mathbf{u} = 0 becomes: u^′TT′−TFT−1u^=0\hat{\mathbf{u}}'^T T'^{-T} F T^{-1} \hat{\mathbf{u}} = 0 which implies that the algebraic matrices are related by F^=T′−TFT−1\hat{F} = T'^{-T} F T^{-1}, or F=T′TF^TF = T'^T \hat{F} T.

    In matrix-vector form, the set of linear equations Af=0A \mathbf{f} = 0 becomes A^f^=0\hat{A} \hat{\mathbf{f}} = 0, where A^=AS\hat{A} = A S for a 9×99 \times 9 transformation matrix SS determined by TT and T′T'. The least-squares solution to Af=0A \mathbf{f} = 0 subject to ∥f∥=1\|\mathbf{f}\| = 1 is the least eigenvector of ATAA^T A, satisfying ATAf=λfA^T A \mathbf{f} = \lambda \mathbf{f}. For the transformed system, multiplying (AS)T(AS)(A S)^T (A S) by S−1fS^{-1} \mathbf{f} gives: STATAS(S−1f)=STATAf=λSTS(S−1f)≠λ(S−1f)S^T A^T A S (S^{-1} \mathbf{f}) = S^T A^T A \mathbf{f} = \lambda S^T S (S^{-1} \mathbf{f}) \neq \lambda (S^{-1} \mathbf{f}) Because STSS^T S is not a scalar multiple of the identity matrix in general, S−1fS^{-1} \mathbf{f} is not an eigenvector of STATASS^T A^T A S. Thus, solving the linear system after coordinate transformation yields a genuinely different fundamental matrix rather than a simple coordinate re-expression.

  5. Knowl 5 — Epipolar Accuracy of Normalized 8-Point Algorithm Compared to Optimal Iterative Estimation

    empirical result

    Across multiple real-world image datasets (including the House, Statue, Museum, Calibration Jig, and Corridor scenes) evaluated over 100 runs across varying subset sizes NN:

    • The unnormalized 8-point algorithm exhibits high and erratic average point-to-epipolar-line distances, frequently ranging from 5 to over 30 pixels on images with realistic noise (e.g., Museum, Statue, and House datasets).
    • The normalized 8-point algorithm achieves stable sub-pixel accuracy, reducing average point-to-epipolar-line errors to between 0.05 and 0.8 pixels across all datasets.
    • The performance of the normalized 8-point algorithm is nearly indistinguishable from the "optimal" iterative algorithm (which minimizes the non-linear reprojection error ∑id(u^i,ui)2+d(u^i′,ui′)2\sum_i d(\hat{\mathbf{u}}_i, \mathbf{u}_i)^2 + d(\hat{\mathbf{u}}'_i, \mathbf{u}'_i)^2 subject to u^i′TFu^i=0\hat{\mathbf{u}}'_i{}^T F \hat{\mathbf{u}}_i = 0 and det⁡F=0\det F = 0), while running approximately 20 times faster.

  6. Knowl 6 — Stage-by-Stage Ablation of Normalization in the 8-Point Algorithm

    empirical result

    Isolating the effects of coordinate normalization across the two stages of the 8-point algorithm—Stage 1 (linear equation solving) and Stage 2 (rank-2 constraint enforcement via SVD)—demonstrates the following error behaviors on real image pairs:

    1. No Normalization (Linear Solution →\rightarrow Rank-2 Enforcement): Average epipolar error remains high and erratic (7–11 pixels for the House dataset across N=8N = 8 to 4040).
    2. Stage 2 Normalization Only (Linear Solution →\rightarrow Normalization →\rightarrow Rank-2 Enforcement →\rightarrow Denormalization): Reduces epipolar error to ≈5\approx 5 pixels, showing that balanced matrix scaling during rank-2 projection provides a notable benefit on its own.
    3. Stage 1 Normalization Only (Normalization →\rightarrow Linear Solution →\rightarrow Denormalization →\rightarrow Rank-2 Enforcement): Reduces error dramatically to ≈1\approx 1 pixel for N>15N > 15, demonstrating that conditioning the linear system provides the largest individual improvement.
    4. Both Stages (Full Normalized Algorithm): Yields the lowest overall error (≈0.4\approx 0.4 pixels for N>15N > 15).

    When N=8N = 8 exactly, Stage 1 normalization has zero impact on error because the system is determined rather than an overdetermined least-squares problem; however, Stage 2 normalization still improves accuracy at N=8N = 8.

  7. Knowl 7 — Sensitivity of Iterative Epipolar Minimization to Algorithm Initialization

    empirical result

    Non-linear iterative methods for computing the fundamental matrix (such as Zhang's gradient-based method or point-epipolar line distance minimization) are sensitive to their starting point:

    • When initialized with the unnormalized 8-point algorithm, iterative algorithms frequently converge to local minima, resulting in higher residual errors than the linear normalized 8-point algorithm (for example, on the House dataset, Zhang's iterative method initialized with unnormalized coordinates achieved an average error of ≈0.6\approx 0.6 pixels at N=30N=30, whereas the normalized 8-point algorithm alone achieved ≈0.45\approx 0.45 pixels).
    • When initialized with the normalized 8-point algorithm, iterative algorithms avoid poor local minima and converge to the optimal solution.

    Providing an accurate, well-conditioned linear initialization via coordinate normalization is more decisive for estimation accuracy than the choice among different non-linear iterative cost functions.

  8. Knowl 8 — 3D Projective Reconstruction Robustness under Coordinate Noise

    empirical result

    In controlled 3D reconstruction experiments using a 128-point ground-truth Euclidean calibration grid with additive zero-mean Gaussian image noise (standard deviation σ∈[0,0.15]\sigma \in [0, 0.15] pixels):

    • Projective reconstruction using the fundamental matrix computed by the unnormalized 8-point algorithm degrades catastrophically with noise, exhibiting average 3D displacement errors of 2.5 to 4.5 units (where 1 unit is the side length of a calibration grid square).
    • Projective reconstruction using the normalized 8-point algorithm degrades smoothly and minimally, maintaining average 3D reconstruction errors below 0.6 units at σ=0.15\sigma = 0.15 pixels.
    • The 3D reconstruction accuracy of the normalized 8-point algorithm is nearly identical to that of the optimal non-linear point displacement minimization algorithm across the entire noise spectrum.

  9. Knowl 9 — Isotropic versus Non-Isotropic Coordinate Normalization

    model/method

    Two data normalization schemes can be applied to point coordinates prior to solving the epipolar linear equations:

    1. Isotropic Scaling: Each image's coordinates are translated so that the centroid of all points lies at the origin (0,0)T(0, 0)^T, and then scaled uniformly by s=2/davgs = \sqrt{2} / d_{\text{avg}}, where davgd_{\text{avg}} is the mean Euclidean distance of the points from the origin. The average normalized point coordinate is (1,1,1)T(1, 1, 1)^T.

    2. Non-Isotropic Scaling: The centroid of points ui=(ui,vi,1)T\mathbf{u}_i = (u_i, v_i, 1)^T is translated to the origin, and coordinates are normalized by an upper-triangular matrix K−1K^{-1} obtained from the Choleski factorization of the second moment matrix: ∑i=1NuiuiT=NKKT\sum_{i=1}^N \mathbf{u}_i \mathbf{u}_i^T = N K K^T Transforming points via u^i=K−1ui\hat{\mathbf{u}}_i = K^{-1} \mathbf{u}_i produces ∑i=1Nu^iu^iT=NI\sum_{i=1}^N \hat{\mathbf{u}}_i \hat{\mathbf{u}}_i^T = N I, forcing both principal moments of the point distribution to equal 1.

    Experimental evaluation demonstrates that the epipolar error curves for isotropic and non-isotropic normalization are virtually indistinguishable. Because isotropic normalization is simpler and computationally lighter, it is the preferred method.

Coverage note — Brief comparative tests with Beardsley and Zisserman's Approximate Calibration and Iterative Linear algorithms were omitted because the paper noted that insufficiently many trials were run on them to reach a firm conclusion.

References

  1. 1.H.C. Longuet-Higgins, “A computer algorithm for reconstructing a scene from two projections,” Nature, vol. 293, pp. 133–135, Sept 1981.
  2. 2.O. D. Faugeras, “What can be seen in three dimensions with an uncalibrated stereo rig?,” in Computer Vision - ECCV ’92, LNCS-Series Vol. 588, Springer-Verlag, 1992, pp. 563 – 578.
  3. 3.R. Hartley, R. Gupta, and T. Chang, “Stereo from uncalibrated cameras,” in Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 1992, pp. 761–764.
  4. 4.Richard Hartley and Rajiv Gupta, “Computing matched-epipolar projections,” in Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 1993, pp. 549 – 555.
  5. 5.Stefan Carlsson, “Multiple image invariants using the double algebra,” in Proc. of the Second Europe-US Workshop on Invariance, Ponta Delgada, Azores, October 1993, pp. 335–350.
  6. 6.R. Deriche, Z. Zhang, Q.-T. Luong, and O. Faugeras, “Robust recovery of the epipolar geometry for an uncalibrated stereo rig,” in Computer Vision - ECCV ’94, Volume I, LNCS-Series Vol. 800, Springer-Verlag, 1994, pp. 567–576.
  7. 7.P. H. S. Torr and D. W. Murray, “Outlier detection and motion segmentation,” in Sensor Fusion VI, P. S. Schenker, Ed. 1993, pp. 432–443, SPIE volume 2059, Boston.
  8. 8.Z. Zhang, R. Deriche, O. Faugeras, and Q.-T. Luong, “A robust technique for matching two uncalibrated images through the recovery of the unknown epipolar geometry,” Artificial Intelligence Journal, vol. 78, pp. 87–119, October 1995.
  9. 9.Quang-Tuan Luong, Rachid Deriche, Olivier D. Faugeras, and Theodore Papadopoulo, “On determining the fundamental matrix: analysis of different methods and experimental results,” Report RR-1894, INRIA, 1993.
  10. 10.Zhengyou Zhang, “Determining the epipolar geometry and its uncertainty : A review,” Report RR-2927, INRIA, 1996.
  11. 11.Richard I. Hartley, “Euclidean reconstruction from uncalibrated views,” in Proc. of the Second Europe-US Workshop on Invariance, Ponta Delgada, Azores, October 1993, pp. 187–202.
  12. 12.P. A. Beardsley, A. Zisserman, and D. W. Murray, “Navigation using affine structure from motion,” in Computer Vision - ECCV ’94, Volume II, LNCS-Series Vol. 801, Springer-Verlag, 1994, pp. 85–96.
  13. 13.S. J. Maybank, “The projective geometry of ambiguous surfaces,” Phil. Trans. R. Soc. Lond., vol. A 332, pp. 1 – 47, 1990.
  14. 14.William H. Press, Brian P. Flannery, Saul A. Teukolsky, and William T. Vetterling, Numerical Recipes in C: The Art of Scientific Computing, Cambridge University Press, 1988.
  15. 15.K.E. Atkinson, An Introduction to Numerical Analysis, 2nd Edition, John Wiley and Sons, New York, 1989.
  16. 16.R. Y. Tsai and T. S. Huang, “Uniqueness and estimation of three dimensional motion parameters of rigid objects with curved surfaces,” IEEE Trans. Patt. Anal. Machine Intell., vol. PAMI-6, pp. 13–27, 1984.
  17. 17.Richard I. Hartley, “Minimizing algebraic distance,” in Proc. DARPA Image Understanding Workshop, 1997.
  18. 18.Fred L. Bookstein, “Fitting conic sections to scattered data,” Computer Graphics and Image Processing, vol. 9, pp. 56 – 71, 1979.
  19. 19.Gene H. Golub and Charles F. Van Loan, Matrix Computations, Second edition, The Johns Hopkins University Press, Baltimore, London, 1989.
  20. 20.R. I. Hartley, “Estimation of relative camera positions for uncalibrated cameras,” in Computer Vision - ECCV ’92, LNCS-Series Vol. 588, Springer-Verlag, 1992, pp. 579 – 587.
  21. 21.Richard I. Hartley, “Projective reconstruction and invariants from multiple images,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 16, pp. 1036–1041, October 1994.
  22. 22.I.E. Sutherland, “Sketchpad: A man-machine graphical communications system,” Technical Report 296, MIT Lincoln Laboratories, 1963, Also published by Garland Publishing Inc, New York, 1980.
  23. 23.Richard I. Hartley, “Projective reconstruction from line correspondences,” in Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 1994, pp. 903–907.
  24. 24.Richard I. Hartley and Peter Sturm, “Triangulation,” in Proc. ARPA Image Understanding Workshop, 1994, pp. 957–966.
  25. 25.R. I. Hartley, “In defence of the 8-point algorithm,” in Proc. International Conference on Computer Vision, 1995, pp. 1064 – 1070.
  26. 26.R. I. Hartley, “A linear method for reconstruction from lies and points,” in Proc. International Conference on Computer Vision, 1995, pp. 882 – 887.
  27. 27.I.E. Sutherland, “Three dimensional data input by tablet,” Proceedings of IEEE, vol. Vol. 62, No. 4, pp. 453–461, April 1974.
  28. 28.B. K. P. Horn, “Relative orientation,” International Journal of Computer Vision, vol. 4, pp. 59 – 78, 1990.

Citation

MLA
Hartley, R. I. “In Defense of the Eight-point Algorithm”. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 19, no. 6, 1997, pp. 580–93, https://doi.org/10.1109/34.601246.
APA
Hartley, R. I. (1997). In defense of the eight-point algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(6), 580–593. https://doi.org/10.1109/34.601246
Chicago
Hartley, R. I. 1997. “In Defense of the Eight-point Algorithm”. IEEE Transactions on Pattern Analysis and Machine Intelligence 19 (6): 580–93. https://doi.org/10.1109/34.601246.
Harvard
Hartley, R.I. (1997) “In defense of the eight-point algorithm”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(6), pp. 580–593. Available at: https://doi.org/10.1109/34.601246.
Vancouver
1. Hartley RI (1997) In defense of the eight-point algorithm. IEEE Transactions on Pattern Analysis and Machine Intelligence 19:580–593

BibTeX

@article{Hartley_1997, title={In defense of the eight-point algorithm}, volume={19}, ISSN={0162-8828}, url={http://dx.doi.org/10.1109/34.601246}, DOI={10.1109/34.601246}, number={6}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Hartley, R.I.}, year={1997}, month=June, pages={580–593} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF