* Multiple View Geometry in Computer Vision, Richard Hartley and Andrew Zisserman, @cite HartleyZ00 (some sample chapters are available [here](https://www.robots.ox.ac.uk/~vgg/hzbook/), CVPR Tutorials are available [here](https://www.robots.ox.ac.uk/~az/tutorials/))
* An Invitation to 3-D Vision: From Images to Geometric Models, Yi Ma, Stefano Soatto, Jana Kosecka, and S. Shankar Sastry, @cite Ma:2003:IVI (a computer vision book handout is available [here](https://cs.gmu.edu/%7Ekosecka/cs685/VisionBookHandout.pdf))
* Computer Vision: Algorithms and Applications, Richard Szeliski, @cite RS10 (an electronic version is available [here](https://szeliski.org/Book/))
* Deeper understanding of the homography decomposition for vision-based control, Ezio Malis, Manuel Vargas, @cite Malis2007 (open access [here](https://hal.inria.fr/inria-00174036))
* Pose Estimation for Augmented Reality: A Hands-On Survey, Eric Marchand, Hideaki Uchiyama, Fabien Spindler, @cite Marchand16 (open access [here](https://hal.inria.fr/hal-01246370))
The homography matrix is a `3x3` matrix but with 8 DoF (degrees of freedom) as it is estimated up to a scale. It is generally normalized (see also \ref lecture_16 "1")
* a rotating camera around its axis of projection, equivalent to consider that the points are on a plane at infinity (image taken from \ref projective_transformations "2")
\note Please note that the code to estimate the camera pose from the homography is an example and you should use instead @ref cv::solvePnP if you want to estimate the camera pose for a planar or an arbitrary object.
The homography can be estimated using for instance the Direct Linear Transform (DLT) algorithm (see \ref lecture_16 "1" for more information).
As the object is planar, the transformation between points expressed in the object frame and projected points into the image plane expressed in the normalized camera frame is a homography. Only because the object is planar,
the camera pose can be retrieved from the homography, assuming the camera intrinsic parameters are known (see \ref projective_transformations "2" or \ref answer_dsp "4").
This can be tested easily using a chessboard object and `findChessboardCorners()` to get the corner locations in the image.
The first thing consists to detect the chessboard corners, the chessboard size (`patternSize`), here `9x6`, is required:
The image points expressed in the normalized camera can be computed from the corner points and by applying a reverse perspective transformation using the camera intrinsics and the distortion coefficients:
This is a quick solution (see also \ref projective_transformations "2") as this does not ensure that the resulting rotation matrix will be orthogonal and the scale is estimated roughly by normalize the first column to 1.
A solution to have a proper rotation matrix (with the properties of a rotation matrix) consists to apply a polar decomposition, or orthogonalization of the rotation matrix
(see \ref polar_decomposition "6" or \ref polar_decomposition_svd "7" or \ref polar_decomposition_svd_2 "8" or \ref Kabsch_algorithm "9" for some information):
In this example, a source image will be transformed into a desired perspective view by computing the homography that maps the source points into the desired points.
The following image shows the source image (left) and the chessboard view that we want to transform into the desired chessboard view (right).
The first step consists to detect the chessboard corners in the source and desired images:
The homography relates the transformation between two planes and it is possible to retrieve the corresponding camera displacement that allows to go from the first to the second plane view (see @cite Malis2007 for more information).
Before going into the details that allow to compute the homography from the camera displacement, some recalls about camera pose and homogeneous transformation.
The function @ref cv::solvePnP allows to compute the camera pose from the correspondences 3D object points (points expressed in the object frame) and the projected 2D image points (object points viewed in the image).
The intrinsic parameters and the distortion coefficients are required (see the camera calibration process).
\f$ \mathbf{K} \f$ is the intrinsic matrix and \f$ ^{c}\mathbf{M}_o \f$ is the camera pose. The output of @ref cv::solvePnP is exactly this: `rvec` is the Rodrigues rotation vector and `tvec` the translation vector.
In this example, we will compute the camera displacement between two camera poses with respect to the chessboard object. The first step consists to compute the camera poses for the two images:
The homography related to a specific plane computed from the camera displacement is:
![By Homography-transl.svg: Per Rosengren derivative work: Appoose (Homography-transl.svg) [CC BY 3.0 (http://creativecommons.org/licenses/by/3.0)], via Wikimedia Commons](images/homography_camera_displacement.png)
On this figure, `n` is the normal vector of the plane and `d` the distance between the camera frame and the plane along the plane normal.
The [equation](https://en.wikipedia.org/wiki/Homography_(computer_vision)#3D_plane_to_plane_equation) to compute the homography from the camera displacement is:
Where \f$ ^{2}\mathbf{H}_{1} \f$ is the homography matrix that maps the points in the first camera frame to the corresponding points in the second camera frame, \f$ ^{2}\mathbf{R}_{1} = \hspace{0.2em} ^{c_2}\mathbf{R}_{o} \cdot \hspace{0.1em} ^{c_1}\mathbf{R}_{o}^{\top} \f$
is the rotation matrix that represents the rotation between the two camera frames and \f$ ^{2}\mathbf{t}_{1} = \hspace{0.2em} ^{c_2}\mathbf{R}_{o} \cdot \left( - \hspace{0.1em} ^{c_1}\mathbf{R}_{o}^{\top} \cdot \hspace{0.1em} ^{c_1}\mathbf{t}_{o} \right ) + \hspace{0.1em} ^{c_2}\mathbf{t}_{o} \f$
the translation vector between the two camera frames.
Here the normal vector `n` is the plane normal expressed in the camera frame 1 and can be computed as the cross product of 2 vectors (using 3 non collinear points that lie on the plane) or in our case directly with:
The distance `d` can be computed as the dot product between the plane normal and a point on the plane or by computing the [plane equation](http://mathworld.wolfram.com/Plane.html) and using the D coefficient:
The projective homography matrix \f$ \textbf{G} \f$ can be computed from the Euclidean homography \f$ \textbf{H} \f$ using the intrinsic matrix \f$ \textbf{K} \f$ (see @cite Malis2007), here assuming the same camera between the two plane views:
The homography matrices are similar. If we compare the image 1 warped using both homography matrices:
Visually, it is hard to distinguish a difference between the result image from the homography computed from the camera displacement and the one estimated with @ref cv::findHomography function.
OpenCV 3 contains the function @ref cv::decomposeHomographyMat which allows to decompose the homography matrix to a set of rotations, translations and plane normals.
First we will decompose the homography matrix computed from the camera displacement:
rvec from homography decomposition: [-0.0919829920641369, -0.5372581036567992, 1.310868863540717]
rvec from camera displacement: [-0.09198299206413783, -0.5372581036567995, 1.310868863540717]
tvec from homography decomposition: [-0.7747961019053186, -0.02751124463434032, -0.6791980037590677] and scaled by d: [-0.1578091561210742, -0.005603443652993778, -0.1383378976078466]
tvec from camera displacement: [0.1578091561210745, 0.005603443652993617, 0.1383378976078466]
plane normal from homography decomposition: [-0.1973513139420648, 0.6283451996579074, -0.7524857267431757]
plane normal at camera 1 pose: [0.1973513139420654, -0.6283451996579068, 0.752485726743176]
Solution 1:
rvec from homography decomposition: [-0.0919829920641369, -0.5372581036567992, 1.310868863540717]
rvec from camera displacement: [-0.09198299206413783, -0.5372581036567995, 1.310868863540717]
tvec from homography decomposition: [0.7747961019053186, 0.02751124463434032, 0.6791980037590677] and scaled by d: [0.1578091561210742, 0.005603443652993778, 0.1383378976078466]
tvec from camera displacement: [0.1578091561210745, 0.005603443652993617, 0.1383378976078466]
plane normal from homography decomposition: [0.1973513139420648, -0.6283451996579074, 0.7524857267431757]
plane normal at camera 1 pose: [0.1973513139420654, -0.6283451996579068, 0.752485726743176]
Solution 2:
rvec from homography decomposition: [0.1053487907109967, -0.1561929144786397, 1.401356552358475]
rvec from camera displacement: [-0.09198299206413783, -0.5372581036567995, 1.310868863540717]
tvec from homography decomposition: [-0.4666552552894618, 0.1050032934770042, -0.913007654671646] and scaled by d: [-0.0950475510338766, 0.02138689274867372, -0.1859598508065552]
tvec from camera displacement: [0.1578091561210745, 0.005603443652993617, 0.1383378976078466]
plane normal from homography decomposition: [-0.3131715472900788, 0.8421206145721947, -0.4390403768225507]
plane normal at camera 1 pose: [0.1973513139420654, -0.6283451996579068, 0.752485726743176]
Solution 3:
rvec from homography decomposition: [0.1053487907109967, -0.1561929144786397, 1.401356552358475]
rvec from camera displacement: [-0.09198299206413783, -0.5372581036567995, 1.310868863540717]
tvec from homography decomposition: [0.4666552552894618, -0.1050032934770042, 0.913007654671646] and scaled by d: [0.0950475510338766, -0.02138689274867372, 0.1859598508065552]
tvec from camera displacement: [0.1578091561210745, 0.005603443652993617, 0.1383378976078466]
plane normal from homography decomposition: [0.3131715472900788, -0.8421206145721947, 0.4390403768225507]
plane normal at camera 1 pose: [0.1973513139420654, -0.6283451996579068, 0.752485726743176]
```
The result of the decomposition of the homography matrix can only be recovered up to a scale factor that corresponds in fact to the distance `d` as the normal is unit length.
As you can see, there is one solution that matches almost perfectly with the computed camera displacement. As stated in the documentation:
```
At least two of the solutions may further be invalidated if point correspondences are available by applying positive depth constraint (all points must be in front of the camera).
As the result of the decomposition is a camera displacement, if we have the initial camera pose \f$ ^{c_1}\mathbf{M}_{o} \f$, we can compute the current camera pose
\f$ ^{c_2}\mathbf{M}_{o} = \hspace{0.2em} ^{c_2}\mathbf{M}_{c_1} \cdot \hspace{0.1em} ^{c_1}\mathbf{M}_{o} \f$ and test if the 3D object points that belong to the plane are projected in front of the camera or not.
Another solution could be to retain the solution with the closest normal if we know the plane normal expressed at the camera 1 pose.
The same thing but with the homography matrix estimated with @ref cv::findHomography
```
Solution 0:
rvec from homography decomposition: [0.1552207729599141, -0.152132696119647, 1.323678695078694]
rvec from camera displacement: [-0.09198299206413783, -0.5372581036567995, 1.310868863540717]
tvec from homography decomposition: [-0.4482361704818117, 0.02485247635491922, -1.034409687207331] and scaled by d: [-0.09129598307571339, 0.005061910238634657, -0.2106868109173855]
tvec from camera displacement: [0.1578091561210745, 0.005603443652993617, 0.1383378976078466]
plane normal from homography decomposition: [-0.1384902722707529, 0.9063331452766947, -0.3992250922214516]
plane normal at camera 1 pose: [0.1973513139420654, -0.6283451996579068, 0.752485726743176]
Solution 1:
rvec from homography decomposition: [0.1552207729599141, -0.152132696119647, 1.323678695078694]
rvec from camera displacement: [-0.09198299206413783, -0.5372581036567995, 1.310868863540717]
tvec from homography decomposition: [0.4482361704818117, -0.02485247635491922, 1.034409687207331] and scaled by d: [0.09129598307571339, -0.005061910238634657, 0.2106868109173855]
tvec from camera displacement: [0.1578091561210745, 0.005603443652993617, 0.1383378976078466]
plane normal from homography decomposition: [0.1384902722707529, -0.9063331452766947, 0.3992250922214516]
plane normal at camera 1 pose: [0.1973513139420654, -0.6283451996579068, 0.752485726743176]
Solution 2:
rvec from homography decomposition: [-0.2886605671759886, -0.521049903923871, 1.381242030882511]
rvec from camera displacement: [-0.09198299206413783, -0.5372581036567995, 1.310868863540717]
tvec from homography decomposition: [-0.8705961357284295, 0.1353018038908477, -0.7037702049789747] and scaled by d: [-0.177321544550518, 0.02755804196893467, -0.1433427218822783]
tvec from camera displacement: [0.1578091561210745, 0.005603443652993617, 0.1383378976078466]
plane normal from homography decomposition: [-0.2284582117722427, 0.6009247303964522, -0.7659610393954643]
plane normal at camera 1 pose: [0.1973513139420654, -0.6283451996579068, 0.752485726743176]
Solution 3:
rvec from homography decomposition: [-0.2886605671759886, -0.521049903923871, 1.381242030882511]
rvec from camera displacement: [-0.09198299206413783, -0.5372581036567995, 1.310868863540717]
tvec from homography decomposition: [0.8705961357284295, -0.1353018038908477, 0.7037702049789747] and scaled by d: [0.177321544550518, -0.02755804196893467, 0.1433427218822783]
tvec from camera displacement: [0.1578091561210745, 0.005603443652993617, 0.1383378976078466]
plane normal from homography decomposition: [0.2284582117722427, -0.6009247303964522, 0.7659610393954643]
plane normal at camera 1 pose: [0.1973513139420654, -0.6283451996579068, 0.752485726743176]
```
Again, there is also a solution that matches with the computed camera displacement.
### Demo 5: Basic panorama stitching from a rotating camera {#tutorial_homography_Demo5}
\note This example is made to illustrate the concept of image stitching based on a pure rotational motion of the camera and should not be used to stitch panorama images.
The [stitching module](@ref stitching) provides a complete pipeline to stitch images.
The homography transformation applies only for planar structure. But in the case of a rotating camera (pure rotation around the camera axis of projection, no translation), an arbitrary world can be considered
The homography can then be computed using the rotation transformation and the camera intrinsic parameters as (see for instance \ref homography_course "10"):
To illustrate, we used Blender, a free and open-source 3D computer graphics software, to generate two camera views with only a rotation transformation between each other.
More information about how to retrieve the camera intrinsic parameters and the `3x4` extrinsic matrix with respect to the world can be found in \ref answer_blender "11" (an additional transformation
* \anchor polar_decomposition_svd 7. [Chapter 3 - 3.1.2 From matrices to rotations - Theorem 3.1 (Least-squares estimation of a rotation from a matrix K)](https://www-sop.inria.fr/asclepios/cours/MVA/Rotations.pdf)
* \anchor polar_decomposition_svd_2 8. [A Personal Interview with the Singular Value Decomposition](https://web.stanford.edu/~gavish/documents/SVD_ans_you.pdf), Matan Gavish
* \anchor Kabsch_algorithm 9. [Kabsch algorithm, Computation of the optimal rotation matrix](https://en.wikipedia.org/wiki/Kabsch_algorithm#Computation_of_the_optimal_rotation_matrix)
* \anchor homography_course 10. [Homography](http://people.scs.carleton.ca/~c_shu/Courses/comp4900d/notes/homography.pdf), Dr. Gerhard Roth
* \anchor answer_blender 11. [3x4 camera matrix from blender camera](https://blender.stackexchange.com/a/38210)
