In this post we will take a look at how Camera Projections work. A demo at the
end will illustrate the important segments of the theory. Prerequisites to
understand the material are available in the *Readings & References* section below.

Epipolar Geometry is the intrinsic projective geometry ^{.css-1i9b1eb{color:var(--theme-ui-colors-primary,#2b6cb0);-webkit-text-decoration:none;text-decoration:none;}.css-1i9b1eb:hover{-webkit-text-decoration:underline;text-decoration:underline;}1} between two views. This
knowledge becomes an interesting piece in the puzzle of estimating the 3D geometry
of a given image projection and the estimated 3D model can then be applied to a
myriad of meaningful real-world problems.

To get there, we will first establish a theoretical framework on how projections are formed and realize some interesting properties.

We will start with a very simple camera where the centre of projection is at the
origin of the Euclidean Coordinate System. Let $Z = f$ be the plane of projection,
more simply the plane where the points in the 3D world space are projected. It is
also called the *image plane* or the *focal plane*.

A Simple Pinhole Camera Model

$C$ is the *camera centre* and $p$ is the *principal point*. By virtue of
similar triangles, we can establish a relation between a point in 3D space at $(X,Y,Z)$
to a point on the focal plane at $(x^\prime,y^\prime)$, a distance of focal length
$f$. This has been visualized in the figure above on the right.

$\frac{f}{Z} = \frac{y}{Y} \leftrightarrow y = \frac{fY}{Z}$

A similar analysis (think of viewing the $ zx $ plane down the $ y $ axis) gives us

$\frac{f}{Z} = \frac{x}{X} \leftrightarrow x = \frac{fX}{Z}$

and we equivalently define a mapping (by dropping the $z$ coordinate)

$(X,Y,Z) \mapsto (x,y) \leftrightarrow ({fX}/{Z},{fY}/{Z})$

More formally, this can be defined in terms of homogeneous coordinates and a matrix transform as

$\begin{pmatrix}X \\ Y \\ Z \\ 1 \end{pmatrix} \mapsto \begin{pmatrix}fX \\ fY \\ Z \end{pmatrix} = \begin{bmatrix}f & 0 & 0 & 0 \\ 0 & f & 0 & 0 \\ 0 & 0 & 1 & 0 \end{bmatrix} \begin{pmatrix}X \\ Y \\ Z \\ 1 \end{pmatrix}$

$x = PX$

This $3 \times 4$ matrix $P$ is known as the **Camera Projection Matrix**.
To adjust for the reference frame in the *image plane* - origin at corner v/s origin at
the center, we add a translation component to the left $3 \times 3$ matrix of $P$.

$K = \begin{bmatrix}f & 0 & p_x \\ 0 & f & p_y \\ 0 & 0 & 1 \end{bmatrix}$

and this is called the **Camera Calibration Matrix**. For added generality to allow for non-square
pixels we multiple the matrix with an extra factor of $diag(m_x,m_y,1)$ where $m$ is the
pixels per unit distance. There's another possibility of distortion due to the $x$ and
$y$ axes not being perpendicular and this is denoted by $s$. The final matrix then
becomes -

$K = \begin{bmatrix}\alpha_x & s & x_0 \\ 0 & \alpha_y & y_0 \\ 0 & 0 & 1 \end{bmatrix}$

In most cases, the camera coordinates will be with respect to a different coordinate system where it is not at the origin in the world coordinate system. And as we had seen in an earlier discussion on projective geometry, Euclidean transformations consist of a rotation and a translation component. Hence, any coordinate can be shifted from the world coordinate space to the camera coordinate space (where the camera is the origin) with the equation -

$X_\text{camera} = R(X_\text{world} - C)$

where $R$ is the rotation matrix and $-C$ represents a translation with respect
to the camera center to make it the origin. Combining the *camera calibration matrix*
and the coordinate frame shifting above, we can draw a mapping between the point in
world coordinate system $X$ and the point $x$ in the image plane, represented by

$x = KR \begin{bmatrix}I \vert -C\end{bmatrix}X$

$P = KR \begin{bmatrix}I \vert -C\end{bmatrix}$

Parameters in $K$ are called the **camera intrinsics** (they never change) and the
remaining parameters $R$ and $C$ are called the **camera extrinsics**.
In literature, these terms will come often and now we know how these are constructed.

Now that we have established all the camera parameters, let us consider two *camera centers*
$C$ and $C^\prime$ with both *image planes* shown in the figure below. These
cameras project a point $ X $ in the world coordinate system to $x$ and $x^\prime$
in respective *image planes*. The plane formed by all these points is known as the **epipolar plane**.

Point Correspondence and Epipolar Geometry

The line joining the two camera centers is known as the **baseline** and intersects both planes at
$e$ and $e^\prime$. These points are known as the **epipoles**, visualized on the right.
We are aiming to establish and constraint between $x$ and $x^\prime$. If one observes
the figure on the right above, the *baseline* ($CC^\prime$) and the ray back-projected from
$x$ ($CX$) form the *epipolar plane* $\pi$. When $\pi$ intersects with the
image plane on the right, it forms the line $l^\prime$ which is known as the **epipolar line**
corresponding to $x$. A search for the point correspondence of $x$ denoted by $x^\prime$
is now constrained to this *epipolar line* which comes as a direct consequence of the way we constructed
$\pi$.

If we rotate the *epipolar plane* with *baseline* as the axis, we get a pencil of planes and all
these planes will still intersect at the *epipoles* $e$ and $e^\prime$. This is a
very interesting result!

We've established some basics for *epipolar geometry* and realized how images are projected
by a given *camera matrix* $ P $. Now we will give a more formal treatment to the problem
both algebraically and geometrically and arrive at a very important matrix known as the
Fundamental Matrix (literally!).

In a previous discussion (see Readings), we had seen **Projectivities** which are mappings
from point in one plane to another. By the above construction of the *epipolar plane*, we
can assert that a similar matrix exists, known as the *Planar Homography* $H_\pi$.

$x^\prime = H_\pi x$

In another interesting result, *Fundamental Matrix* does not need the *epipolar plane* to be
necessarily defined. Let us see how it directly relates to the *camera matrices* we described
above.

If one remembers the alternate definition of *epiline* corresponding to point $x$, it was
the image of the ray back-projected to the world coordinate $X$ in the second camera view.
The solution to the equation

$x = PX$

is given by a family of solutions in the form of

$X(\lambda) = P^+x + \lambda C$

To find the *epipolar line* $l^\prime$ for point $x$, we know from the above
discussion that both the *epipole* $e^\prime$ and the point correspondence $x^\prime$
lie on the *epiline*. The intersection of these (from our *projective geometry primer*) gives us the line,

$l^\prime = e^\prime \times x^\prime = [e^\prime]_\times x$

$l^\prime = [e^\prime]_\times H_\pi x = Fx$

where $F = [e^\prime]_\times H_\pi$ is known as the **Fundamental Matrix**, a matrix
of rank 2. In plain words, $F$ represents a mapping of the projective plane $\mathbb{P}^2$
of the first image to a pencil of *epipolar lines* $\mathbb{P}^1$ through $e^\prime$.

Note that the notation $[e]_\times$ is a skew-symmetric matrix equivalent for the cross product.

$[e]_\times = \begin{bmatrix} 0 & -e_3 & e_2 \\ e_3 & 0 & -e_1 \\ -e_2 & e_1 & 0 \end{bmatrix}$

where $P^+$ is the pseudo-inverse of $P$, giving $PP^+ = I$ and
$C$ is the right-null space of $ P $ commonly called the *camera center*.
This has its own interesting little derivation and out of scope here.

We have two known points $C$ and $P^+x$ on the ray which can be projected
onto the second image plane via a second camera matrix $P^\prime$ and the *epiline*
is the cross-product of these two projected points.

$l^\prime = (P^\prime C) \times (P^\prime P^+ x)$

We have just arrived at an alternate representation of $l^\prime = e^\prime \times x^\prime$
and hence we realize that *epipole* $e^\prime$ is actually the image of the other
*camera center*. Again we see that

$l^\prime = [e^\prime]_\times (P^\prime P^+ x) = Fx$

where $F = [e^\prime]_\times P^\prime P^+$ and $H_\pi = P^\prime P^+$.
Wow! We've discovered quite a few important relations between *Homography*,
*Camera Matrices* and the *Fundamental Matrix*.

After all the details above, it is easy to arrive and prove at the following result

$x^{\prime T}Fx = 0$

where $x \leftrightarrow x^\prime$ are point-to-point correspondences in the two image planes.
Solution to the above equation is a fairly popular research problem. For all practical
purposes, the above problem generally remains over-determined with an exact solution determined
by 3 non-collinear point correspondences. Hence, techniques like the *family of RANSAC* are used
to non-deterministically determine the best fit model via a relevant cost function like the
re-projection error.

Each execution requires a stereo image pair (for this purpose I shot two-view
images around my house) which represent the two image planes discussed above.
To arrive at $x\leftrightarrow x^\prime$ point correspondences, I use *SURF* to
detect keypoints in both images, match them via *FLANN* and apply *Lowe's ratio test* to
filter out matches that are too far away and might be false positives. If all above
sounds alien, consider it as a black box as it is out of scope for this discussion.

Then we estimate the **Fundamental Matrix** via *RANSAC* ^{2}. One we get that we draw the
*epilines* on both images via $l = F^T x^\prime$ for the left *image plane* and
$l^\prime = Fx$ for the right *image plane*. These line equations are a direct
consequence of the result $x^{\prime T}Fx = 0$.

Full code is available upon request but for the purpose of this discussion, the one below should be easy to follow.

### Load the stereo image pair#image_l = cv2.pyrDown(cv2.imread('images/view_left.jpg', 0))image_r = cv2.pyrDown(cv2.imread('images/view_right.jpg', 0))### Consider this matcher a black box for the purpose of this discussion#matcher = SURFKeyPointMatcher()kp_l, des_l, kp_r, des_r = matcher.detect_and_compute(image_l, image_r)pts_l, pts_r, good_matches = matcher.find_good_matches(kp_l, des_l, kp_r, des_r)### Estimate the Fundamental Matrix via RANSAC#pts_l = np.int32(pts_l)pts_r = np.int32(pts_r)F, mask = cv2.findFundamentalMat(pts_l, pts_r, cv2.FM_RANSAC)### Use only the points which RANSAC determined to be model inliers#pts_l = pts_l[mask.ravel() == 1]pts_r = pts_r[mask.ravel() == 1]### Compute epilines. Note that right image points are used for left view# and left image points are used for right view#epilines_l = cv2.computeCorrespondEpilines(pts_r, 2, F).reshape(-1, 3)epilines_r = cv2.computeCorrespondEpilines(pts_l, 1, F).reshape(-1, 3)image_epilines_l = draw_epilines(image_l, epilines_l)image_epilines_r = draw_epilines(image_r, epilines_r)cv2.imwrite('epi_left.jpg', image_epilines_l)cv2.imwrite('epi_left.jpg', image_epilines_r)

Take a look at the following resulting images from the above code which show a
general sense of direction of both the camera centers. The *epipoles*
(the point of intersection of the *epilines*) lie outside the visible *image planes*.

Left View Epilines (Chair)

Right View Epilines (Chair)

Left View Epilines (Room)

Right View Epilines (Room)

It goes without saying that the epilines constructed are only as good as the $x \leftrightarrow x^\prime$ point correspondences. The robustness of the point correspondences is dependent on how well our keypoint detection algorithm works.

- Hartley, R. & Zisserman, A., 2003. Multiple view geometry in computer vision, Cambridge university press.

- See A Primer on Projective Geometry.↩
- RANSAC was covered in a previous post↩

Created: Tue Aug 08 2017