Automatic Panoramic Image Stitching using Invariant Features

Matthew A. BrownD. Lowe

article2007IJCV2,890 citations

Presents a fully automated system for recognizing and stitching multiple panoramas from unordered, uncalibrated image collections by combining scale-invariant feature matching, probabilistic verification, bundle adjustment, and multi-band blending without requiring manual initialization.

Listen

This paper presents a fully automated method for creating panoramic images from unordered collections of photographs. The core challenge is that existing stitching techniques either require user guidance to align images or assume a fixed sequence and limited variation in camera motion, scale, or lighting, which limits their use on casual photo sets. The authors address this by treating panorama construction as a multi-image matching problem solved through invariant local features rather than direct pixel comparison or manual initialization.

The approach begins by extracting SIFT features from every image and matching them efficiently with a k-d tree. Candidate overlaps are identified by selecting the strongest feature matches, then refined with RANSAC to estimate homographies and a probabilistic test to confirm genuine image pairs. Connected components of verified matches define individual panoramas, after which bundle adjustment jointly optimizes all camera rotations and focal lengths. Additional steps compensate for brightness differences between images, straighten the final mosaic using a horizon heuristic, and apply multi-band blending to hide seams while preserving detail. The method was demonstrated on collections containing dozens of images, including cases with zoom, flash, and varying exposure.

The system reliably detects multiple panoramas within noisy image sets and rejects unrelated photos. It produces seamless 360-degree results even when input images differ substantially in orientation, scale, and illumination. On a 57-image test sequence the entire registration completed in about one minute and yielded a 23-megapixel output without visible artifacts from gain variation. These capabilities remove the need for controlled capture or manual alignment, making high-quality panorama creation practical for everyday photography.

The results matter because they convert an error-prone, interactive process into a robust, one-click operation that works on arbitrary photo collections. This lowers barriers for applications in documentation, virtual tourism, and image archiving. The method also shows that feature-based matching combined with global optimization can handle the geometric and photometric inconsistencies typical of handheld photography.

Future improvements should incorporate radial distortion parameters into bundle adjustment and rendering, model small camera translations to reduce parallax, and detect moving objects so that seams can avoid them. Additional photometric calibration for vignetting and high-dynamic-range fusion would further reduce artifacts in challenging scenes. The current implementation already demonstrates that these extensions are feasible once the core matching and alignment pipeline is in place.

University of British Columbia high resolution image image matching image sequence Invariant Feature SIFT RANSAC

Brown et al (2007).pdf

Paper: Distinctive Image Features from Scale-Invariant Keypoints, David G. Lowe (2004). Lowe's SIFT descriptor provides the foundational invariant local feature extraction mechanism that the source paper relies on for robust multi-image matching.

Paper: Structure-from-Motion Revisited, Johannes L. Schönberger et al. (2016). This paper extends foundational image stitching and matching techniques into modern, robust Structure-from-Motion pipelines that handle unstructured Internet photo collections at scale.

Cover for Automatic Panoramic Image Stitching using Invariant Features

Abstract

This paper concerns the problem of fully automated panoramic image stitching. Though the 1D problem (single axis of rotation) is well studied, 2D or multi-row stitching is more difficult. Previous approaches have used human input or restrictions on the image sequence in order to establish matching images. In this work, we formulate stitching as a multi-image matching problem, and use invariant local features to find matches between all of the images. Because of this our method is insensitive to the ordering, orientation, scale and illumination of the input images. It is also insensitive to noise images that are not part of a panorama, and can recognise multiple panoramas in an unordered image dataset. In addition to providing more detail, this paper extends our previous work in the area [BL03] by introducing gain compensation and automatic straightening steps.

Topics

high resolution image image matching image sequence local feature single-axis

Keywords

Invariant Feature SIFT RANSAC Bundle Adjustment Multi-Band Blending Gain Compensation Panorama Straightening Homography k-d Tree Probabilistic Model

2 Feature Matching
1 Introduction
3 Image Matching
3.1 Robust Homography Estimation using RANSAC
3.2 Probabilistic Model for Image Match Verification
4 Bundle Adjustment
4.1 Fast Solution by Direct Computation of the Linear System
5 Automatic Panorama Straightening
6 Gain Compensation
7 Multi-Band Blending
8 Results
9 Conclusions
Future Work
References

Knowls

Knowl 1 — Fully Automated Panoramic Image Stitching Pipeline
algorithm
The complete panoramic stitching algorithm takes an unordered collection of images, automatically detects panoramas and rejects noise images, aligns the geometry, corrects exposures, and renders seamless composite mosaics.
Input: An unordered set of $n$ images
Output: One or more blended panoramic images
1. Extract SIFT keypoints and invariant feature descriptors from all $n$ images.
2. Match each feature to its $k=4$ nearest neighbours across all images using a k-d tree.
3. For each image $i$ :
    a. Select $m=6$ candidate matching images having the largest number of feature matches to image $i$ .
    b. For each candidate pair $(i, j)$ , robustly estimate homography $H_{ij}$ using RANSAC with 4-point Direct Linear Transformation ( $500$ trials).
    c. Verify whether image pair $(i, j)$ is a valid match using the probabilistic verification condition $n_i > 8.0 + 0.3 n_f$ , where $n_f$ is overlapping feature count and $n_i$ is inlier count.
4. Form an undirected match graph where vertices are images and edges are verified pairs. Find connected components; discard components of size 1 as noise images.
5. For each connected component containing 2 or more images:
    a. Jointly optimize all camera rotation vectors $\theta_i$ and focal lengths $f_i$ using sparse Levenberg-Marquardt bundle adjustment with Huber robust loss.
    b. Apply automatic straightening by aligning the horizon up-vector with the vertical rendering axis.
    c. Estimate per-image photometric gains $g_i$ by solving the closed-form quadratic intensity consistency system.
    d. Render the final panorama in spherical coordinates using multi-band frequency blending.
Knowl 2 — Probabilistic Model for Image Match Verification
model/method
To decide whether two candidate images actually overlap or represent a false match, a Bayesian verification test evaluates the ratio of inlier to outlier feature matches after RANSAC homography estimation.

Let $n_f$ be the total number of features in the geometric overlap region between two images, and let $n_i$ be the number of RANSAC inliers. The match state is a binary variable $m \in \{0, 1\}$ , where $m=1$ indicates a true match and $m=0$ indicates a false match. Each feature correspondence match indicator $f^{(i)} \in \{0, 1\}$ is modeled as an independent Bernoulli trial, giving a Binomial distribution over $n_i$ :

$p(f^{(1:n_f)} \mid m=1) = B(n_i; n_f, p_1)$

$p(f^{(1:n_f)} \mid m=0) = B(n_i; n_f, p_0)$

where $B(x; n, p) = \frac{n!}{x!(n-x)!} p^x (1-p)^{n-x}$ , $p_1 = 0.6$ is the inlier probability given a correct match, and $p_0 = 0.1$ is the inlier probability given a false match.

Applying Bayes' rule, an image match is accepted if the posterior probability $p(m=1 \mid f^{(1:n_f)}) > p_{\min}$ :

$\frac{B(n_i; n_f, p_1) p(m=1)}{B(n_i; n_f, p_0) p(m=0)} > \frac{1}{\frac{1}{p_{\min}} - 1}$

Setting prior $p(m=1) = 10^{-6}$ , $p(m=0) = 1 - 10^{-6}$ , and threshold $p_{\min} = 0.999$ yields the linear decision boundary:

$n_i > \alpha + \beta n_f$

with parameters $\alpha = 8.0$ and $\beta = 0.3$ .
Knowl 3 — Multi-Image Bundle Adjustment with Huber Robust Loss
model/method
Bundle adjustment optimizes all camera parameters jointly to prevent accumulated drift and ensure panoramic loops close seamlessly. Each camera $i$ has 4 parameters $\Phi_i = (\theta_{i1}, \theta_{i2}, \theta_{i3}, f_i)$ , where $\theta_i$ is a 3D rotation vector defining rotation matrix $R_i = e^{[\theta_i]_\times}$ and $f_i$ is the focal length defining calibration matrix $K_i = \text{diag}(f_i, f_i, 1)$ .

For a feature correspondence $u_i^k \leftrightarrow u_j^l$ between images $i$ and $j$ , the homogeneous coordinate prediction of feature $l$ from image $j$ projected into image $i$ is:

$\tilde{p}_{ij}^k = K_i R_i R_j^T K_j^{-1} \tilde{u}_j^l$

giving residual $r_{ij}^k = u_i^k - p_{ij}^k$ .

The total robustified projection error minimized over all images is:

$e = \sum_{i=1}^n \sum_{j \in \mathcal{I}(i)} \sum_{k \in \mathcal{F}(i,j)} h(r_{ij}^k)$

where $\mathcal{I}(i)$ is the set of images matching image $i$ , $\mathcal{F}(i, j)$ is the set of feature matches between $i$ and $j$ , and $h(x)$ is the Huber robust loss function:

$h(x) = \begin{cases} |x|^2, & \text{if } |x| < \sigma \\ 2\sigma|x| - \sigma^2, & \text{if } |x| \ge \sigma \end{cases}$

with $\sigma = \infty$ during image addition initialization and $\sigma = 2\text{ pixels}$ for the final non-linear solve.

Optimization proceeds via Levenberg-Marquardt updates:

$\Delta \Phi = (J^T J + \lambda C_p^{-1})^{-1} J^T r$

where $C_p$ is a diagonal prior covariance matrix on the parameter updates with rotation standard deviation $\sigma_\theta = \pi/16\text{ rad}$ and focal length standard deviation $\sigma_f = \bar{f}/10$ , where $\bar{f}$ is the mean focal length estimated up to that step.
Knowl 4 — Multi-Band Blending for Multi-Image Spherical Mosaics
algorithm
Multi-band blending prevents high-frequency ghosting and edge artifacts caused by parallax, motion, or slight misregistrations, while smoothly equalizing low-frequency illumination variations across $n$ overlapping images.
Input: Set of $n$ registered images $I^i(\theta, \phi)$ in spherical coordinates and number of bands
Output: Blended composite spherical panorama $I^{\text{multi}}(\theta, \phi)$
1. Compute per-image spatial weights $W(x, y) = w(x)w(y)$ where $w(x)$ decreases linearly from $1$ at the image center to $0$ at the image boundary. Resample into spherical weights $W^i(\theta, \phi)$ .
2. Initialize binary max-weight maps for each image $i$ :
    $W_{\max}^i(\theta, \phi) = 1$ if $W^i(\theta, \phi) = \arg\max_j W^j(\theta, \phi)$ , else $0$ .
3. Form the highest-frequency band ( $k=0$ ):
   a. Low-pass image: $I_\sigma^i(\theta, \phi) = I^i(\theta, \phi) * g_\sigma(\theta, \phi)$
   b. Bandpass image: $B_\sigma^i(\theta, \phi) = I^i(\theta, \phi) - I_\sigma^i(\theta, \phi)$
   c. Blending weight: $W_\sigma^i(\theta, \phi) = W_{\max}^i(\theta, \phi) * g_\sigma(\theta, \phi)$
   where $g_\sigma$ is a Gaussian blurring kernel of standard deviation $\sigma$ .
4. For each subsequent lower frequency band $k \ge 1$ :
   a. Blurring kernel standard deviation: $\sigma' = \sqrt{2k+1}\,\sigma$
   b. Next low-pass image: $I_{(k+1)\sigma}^i = I_{k\sigma}^i * g_{\sigma'}$
   c. Bandpass image: $B_{(k+1)\sigma}^i = I_{k\sigma}^i - I_{(k+1)\sigma}^i$
   d. Blending weight: $W_{(k+1)\sigma}^i = W_{k\sigma}^i * g_{\sigma'}$
5. For each band $k$ , combine images linearly using the blurred weights:
    $I_{k\sigma}^{\text{multi}}(\theta, \phi) = \frac{\sum_{i=1}^n B_{k\sigma}^i(\theta, \phi) W_{k\sigma}^i(\theta, \phi)}{\sum_{i=1}^n W_{k\sigma}^i(\theta, \phi)}$
6. Sum all combined frequency band images $I_{k\sigma}^{\text{multi}}(\theta, \phi)$ together with the lowest-frequency residual image to produce the output panorama.
Knowl 5 — Automatic Panorama Straightening via Camera Axis Covariance
model/method
Solving bundle adjustment determines relative camera rotations up to an arbitrary global 3D rotation, which frequently results in a curved or wavy horizon in output panoramas.

Assuming camera operators rarely introduce roll (twist) relative to the horizon, the camera coordinate $X$ -axes (horizontal vectors in each camera's local coordinate frame) lie approximately in a single plane perpendicular to gravity in world space.

Let $X_i$ denote the local horizontal unit vector of camera $i$ transformed into the global coordinate frame. The world up-vector $u$ (normal to the horizontal plane) satisfies:

$\left( \sum_{i=1}^n X_i X_i^T \right) u = 0$

The up-vector $u$ is found by computing the eigenvector corresponding to the smallest eigenvalue of the covariance matrix $\sum_{i=1}^n X_i X_i^T$ . Applying a global rotation to all cameras that maps $u$ to the vertical rendering axis $[0, 1, 0]^T$ eliminates the wavy artifact and straightens the panorama automatically.
Knowl 6 — Photometric Gain Compensation via Closed-Form Quadratic Optimization
model/method
Variations in camera exposure, aperture, and lighting between images are balanced by estimating a scalar gain parameter $g_i$ for each image $i \in \{1, \dots, n\}$ .

The gain error function penalizes differences in gain-adjusted mean intensities over pairwise overlap regions $R(i, j)$ and includes a regularization prior penalizing deviation from unit gain ( $g_i = 1$ ):

$e = \frac{1}{2} \sum_{i=1}^n \sum_{j=1}^n N_{ij} \left( \frac{(g_i \bar{I}_{ij} - g_j \bar{I}_{ji})^2}{\sigma_N^2} + \frac{(1 - g_i)^2}{\sigma_g^2} \right)$

where $N_{ij} = |R(i, j)|$ is the number of pixels in image $i$ that overlap image $j$ , $\bar{I}_{ij} = \frac{1}{N_{ij}} \sum_{u_i \in R(i, j)} I_i(u_i)$ is the mean intensity of image $i$ inside the overlap region with image $j$ , $\sigma_N = 10.0$ is the standard deviation of normalized intensity errors (for pixel values in $[0, 255]$ ), and $\sigma_g = 0.1$ is the standard deviation of the unit-gain prior.

Because $e$ is strictly quadratic in the vector of gains $g = [g_1, g_2, \dots, g_n]^T$ , setting $\frac{\partial e}{\partial g_i} = 0$ for all $i$ produces a symmetric linear system of size $n \times n$ that is solved directly in closed form.
Knowl 7 — Direct Computation of Normal Equations in Sparse Bundle Adjustment
theoretical result
In panoramic bundle adjustment with $N$ camera parameters ( $N = 4n$ for $n$ images) and $M$ residual measurements, forming the normal equations matrix $J^T J$ via explicit dense multiplication of the Jacobian $J$ requires $O(M N^2)$ operations.

Because each image matches only a small local subset of images, the Jacobian $J = \frac{\partial r}{\partial \Phi}$ is block sparse. The non-zero blocks of $J^T J$ and $J^T r$ can be computed by summing directly over observed feature correspondences between overlapping pairs $(i, j)$ :

$(J^T J)_{ij} = \sum_{k \in \mathcal{F}(i,j)} \left( \frac{\partial r_{ij}^k}{\partial \Phi_i} \right)^T \left( \frac{\partial r_{ij}^k}{\partial \Phi_j} \right)$

$(J^T r)_i = \sum_{j \in \mathcal{I}(i)} \sum_{k \in \mathcal{F}(i,j)} \left( \frac{\partial r_{ij}^k}{\partial \Phi_i} \right)^T r_{ij}^k$

where $\mathcal{I}(i)$ is the set of images overlapping image $i$ and $\mathcal{F}(i, j)$ is the set of feature matches between images $i$ and $j$ .

This direct accumulation avoids the $O(M N^2)$ dense matrix product, reducing the per-iteration cost of bundle adjustment to $O(N^3)$ , which is the cost of solving the $N \times N$ linear system.
Knowl 8 — Four-Parameter Camera Model and Pairwise Homographies
model/method
Assuming the camera rotates purely around its optical center, the transformation between any two panoramic views is described by a homography parameterized by 4 extrinsic and intrinsic parameters per image.

Each camera $i$ is parameterized by a 3-element rotation vector $\theta_i = [\theta_{i1}, \theta_{i2}, \theta_{i3}]^T$ and a scalar focal length $f_i$ . The intrinsic calibration matrix $K_i$ is:

$K_i = \begin{bmatrix} f_i & 0 & 0 \\ 0 & f_i & 0 \\ 0 & 0 & 1 \end{bmatrix}$

The 3D rotation matrix $R_i$ is defined via the exponential map of the skew-symmetric cross-product matrix $[ heta_i]_\times$ :

$R_i = e^{[\theta_i]_\times}, \quad [\theta_i]_\times = \begin{bmatrix} 0 & -\theta_{i3} & \theta_{i2} \\ \theta_{i3} & 0 & -\theta_{i1} \\ -\theta_{i2} & \theta_{i1} & 0 \end{bmatrix}$

The pairwise homography $H_{ij}$ mapping homogeneous coordinates $\tilde{u}_j = s_j [u_j, 1]^T$ in image $j$ to homogeneous coordinates $\tilde{u}_i$ in image $i$ is:

$H_{ij} = K_i R_i R_j^T K_j^{-1}$

Local linearized patches around image coordinates undergo an affine transformation $A_{ij} = \frac{\partial u_i}{\partial u_j}$ , which justifies the use of scale- and orientation-invariant local features that are partially robust to affine transformations.
Knowl 9 — Robust Pairwise Homography Estimation using RANSAC
algorithm
To find valid geometric alignments between candidate matching image pairs, RANSAC is applied to SIFT feature correspondences.
Input: Candidate pair of images $(i, j)$ with a set of putative SIFT feature correspondences, number of trials $n=500$ , inlier distance threshold $\epsilon$
Output: Best homography $H_{ij}$ and set of inlier correspondences
1. For trial $t = 1$ to $n=500$ :
    a. Randomly select a minimal sample of $r = 4$ feature correspondences between image $i$ and image $j$ .
    b. Compute candidate homography $H$ using the Direct Linear Transformation (DLT) algorithm.
    c. Project all feature correspondences using $H$ and count the number of inliers whose Euclidean projection error is less than $\epsilon$ pixels.
2. Select the homography $H_{ij}$ that maximizes the number of inliers.
Given the inlier probability $p_i$ that a feature correspondence is correct, the probability of finding the correct transformation after $n$ trials is:

$p(H \text{ is correct}) = 1 - (1 - (p_i)^r)^n$

For an inlier probability $p_i = 0.5$ and $r = 4$ , 500 trials yields a failure probability of approximately $1 \times 10^{-14}$ .
Knowl 10 — Robustness of Invariant Feature Panorama Stitching to Radial Distortion
empirical result
Evaluating panorama stitching on a 44-image test sequence subjected to synthetic first-order radial distortion $x' = (1 + \kappa |x|^2)x$ across $\kappa \in [-0.5, 0.5]$ (with image height normalized to 1 unit):
1. Invariant feature matching and RANSAC homography estimation remain robust without explicitly modeling radial distortion; in the worst-case distortion ( $\kappa = \pm 0.5$ ), the number of consistent matches per feature decreases by approximately one-third, but leaves approximately 500 correct matches per image, sufficient to reliably identify and match images.
2. However, unmodelled radial distortion causes noticeable visual misregistration artifacts in the final rendered output panorama, demonstrating that radial distortion parameters must be included in bundle adjustment and rendering stages for high-quality mosaic generation.

Coverage note — Omitted standard mathematical derivations of the Direct Linear Transformation (DLT), standard SIFT descriptor construction details, and generic Levenberg-Marquardt step equations, as they are standard background techniques.

References

1.A. Agarwala, M. Dontcheva, M. Agarwala, S. Drucker, A. Colburn, B. Curless, D. Salesin, and M. Cohen. Interactive digital photomontage. In ACM Transactions on Graphics (SIGGRAPH’04), 2004.
2.P. Burt and E. Adelson. A multiresolution spline with application to image mosaics. ACM Transactions on Graphics, 2(4):217–236, 1983.
3.B. Bascle, A. Blake, and A. Zisserman. Motion deblurring and super-resolution from and image sequence. In Proceedings of the 4th European Conference on Computer Vision (ECCV96), pages 312–320. Springer-Verlag, 1996.
4.J. Beis and D. Lowe. Shape indexing using approximate nearest-neighbor search in high-dimensional spaces. In Proceedings of the Interational Conference on Computer Vision and Pattern Recognition (CVPR97), pages 1000–1006, 1997.
5.M. Brown and D. Lowe. Recognising panoramas. In Proceedings of the 9th International Conference on Computer Vision (ICCV03), volume 2, pages 1218–1225, Nice, October 2003.
6.D. Brown. Close-range camera calibration. Photogrammetric Engineering, 37(8):855–866, 1971.
7.M. Brown, R. Szeliski, and S. Winder. Multi-image matching using multi-scale oriented patches. In Proceedings of the Interational Conference on Computer Vision and Pattern Recognition (CVPR05), San Diego, June 2005.
8.S. Chen. QuickTime VR – An image-based approach to virtual environment navigation. In SIGGRAPH’95, volume 29, pages 29–38, 1995.
9.D. Capel and A. Zisserman. Automated mosaicing with super-resolution zoom. In Proceedings of the Interational Conference on Computer Vision and Pattern Recognition (CVPR98), pages 885–891, June 1998.
10.J. Davis. Mosaics of scenes with moving objects. In Proceedings of the Interational Conference on Computer Vision and Pattern Recognition (CVPR98), pages 354–360, 1998.
11.P. Debevec and J. Malik. Recovering high dynamic range radiance maps from photographs. Computer Graphics, 31:369–378, 1997.
12.M. Fischler and R. Bolles. Random sample consensus: A paradigm for model fitting with application to image analysis and automated cartography. Communications of the ACM, 24:381–395, 1981.
13.D.B. Goldman and J.H. Chen. Vignette and exposure calibation and compensation. In Proceedings of the 10th International Conference on Computer Vision (ICCV05), pages I: 899–906, 2005.
14.C. Harris. Geometry from visual motion. In A. Blake and A. Yuille, editors, Active Vision, pages 263–284. MIT Press, 1992.
15.P.J. Huber. Robust Statistics. Wiley, 1981.
16.R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004.
17.M. Irani and P. Anandan. About direct methods. In B. Triggs, A. Zisserman, and R. Szeliski, editors, Vision Algorithms: Theory and Practice, number 1883 in LNCS, pages 267–277. Springer-Verlag, Corfu, Greece, September 1999.
18.D. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91–110, 2004.
19.J. Meehan. Panoramic Photography. Amphoto Books, September 1990.
20.D. Milgram. Computer methods for creating photomosaics. IEEE Transactions on Computers, C-24(11):1113–1119, November 1975.
21.P. McLauchlan and A. Jaenicke. Image mosaicing using sequential bundle adjustment. Image and Vision Computing, 20(9-10):751–759, August 2002.
22.Microsoft Digital Image Pro. http://www.microsoft.com/products/imaging.
23.C. Rother and S. Carlsson. Linear multi view reconstruction and camera recovery using a reference plane. International Journal of Computer Vision, 49(2/3):117–141, 2002.
24.Realviz. http://www.realviz.com.
25.H. Seetzen, W. Heidrich, W. Stuerzlinger, G. Ward, L. Whitehead, M. Trentacoste, A. Ghosh, and A. Vorozcovs. High dynamic range display systems. In ACM Transactions on Graphics (SIGGRAPH’04), 2004.
26.R. Szeliski and S. Kang. Direct methods for visual scene reconstruction. In IEEE Workshop on Representations of Visual Scenes, pages 26–33, Cambridge, MA, 1995.
27.H. Sawhney and R. Kumar. True multi-image alignment and its application to mosaicing and lens distortion correction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(3):235–243, 1999.
28.R. Szeliski and H. Shum. Creating full view panoramic image mosaics and environment maps. Computer Graphics (SIGGRAPH’97), 31(Annual Conference Series):251–258, 1997.
29.H. Shum and R. Szeliski. Construction of panoramic mosaics with global and local alignment. International Journal of Computer Vision, 36(2):101–130, February 2000.
30.J. Shi and C. Tomasi. Good features to track. In Proceedings of the Interational Conference on Computer Vision and Pattern Recognition (CVPR94), Seattle, June 1994.
31.J. Sivic and A. Zisserman. Video Google: A text retrieval approach to object matching in videos. In Proceedings of the 9th International Conference on Computer Vision (ICCV03), October 2003.
32.R. Szeliski. Image alignment and stitching: A tutorial. Technical Report MSR-TR-2004-92, Microsoft Research, December 2004.
33.W. Triggs, P. McLauchlan, R. Hartley, and A. Fitzgibbon. Bundle adjustment: A modern synthesis. In Vision Algorithms: Theory and Practice, number 1883 in LNCS, pages 298–373. Springer-Verlag, Corfu, Greece, September 1999.
34.P. Torr. Bayesian model estimation and selection for epipolar geometry and generic manifold fitting. International Journal of Computer Vision, 50(1):35–61, 2002.
35.M. Uyttendaele, A. Eden, and R. Szeliski. Eliminating ghosting and exposure artifacts in image mosaics. In Proceedings of the Interational Conference on Computer Vision and Pattern Recognition (CVPR01), volume 2, pages 509–516, Kauai, Hawaii, December 2001.
36.I. Zoghlami, O. Faugeras, and R. Deriche. Using geometric corners to build a 2D mosaic from a set of images. In Proceedings of the International Conference on Computer Vision and Pattern Recognition, Puerto Rico. IEEE, June 1997.

Citation

MLA

Brown, M., and D. G. Lowe. “Automatic Panoramic Image Stitching Using Invariant Features”. International Journal of Computer Vision, vol. 74, no. 1, 2006, pp. 59–73, https://doi.org/10.1007/s11263-006-0002-3.

APA

Brown, M., & Lowe, D. G. (2006). Automatic Panoramic Image Stitching using Invariant Features. International Journal of Computer Vision, 74(1), 59–73. https://doi.org/10.1007/s11263-006-0002-3

Chicago

Brown, M., and D. G. Lowe. 2006. “Automatic Panoramic Image Stitching Using Invariant Features”. International Journal of Computer Vision 74 (1): 59–73. https://doi.org/10.1007/s11263-006-0002-3.

Harvard

Brown, M. and Lowe, D.G. (2006) “Automatic Panoramic Image Stitching using Invariant Features”, International Journal of Computer Vision, 74(1), pp. 59–73. Available at: https://doi.org/10.1007/s11263-006-0002-3.

Vancouver

1. Brown M, Lowe DG (2006) Automatic Panoramic Image Stitching using Invariant Features. International Journal of Computer Vision 74:59–73

BibTeX

@article{Brown_2006, title={Automatic Panoramic Image Stitching using Invariant Features}, volume={74}, ISSN={1573-1405}, url={http://dx.doi.org/10.1007/s11263-006-0002-3}, DOI={10.1007/s11263-006-0002-3}, number={1}, journal={International Journal of Computer Vision}, publisher={Springer Science and Business Media LLC}, author={Brown, Matthew and Lowe, David G.}, year={2006}, month=Dec, pages={59–73} }

Metadata:Crossref

Access the Paper

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

Open PDF

Automatic Panoramic Image Stitching using Invariant Features

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Fully Automated Panoramic Image Stitching Pipeline

Knowl 2 — Probabilistic Model for Image Match Verification

Knowl 3 — Multi-Image Bundle Adjustment with Huber Robust Loss

Knowl 4 — Multi-Band Blending for Multi-Image Spherical Mosaics

Knowl 5 — Automatic Panorama Straightening via Camera Axis Covariance

Knowl 6 — Photometric Gain Compensation via Closed-Form Quadratic Optimization

Knowl 7 — Direct Computation of Normal Equations in Sparse Bundle Adjustment

Knowl 8 — Four-Parameter Camera Model and Pairwise Homographies

Knowl 9 — Robust Pairwise Homography Estimation using RANSAC

Knowl 10 — Robustness of Invariant Feature Panorama Stitching to Radial Distortion

References

Citation

BibTeX

Access the Paper