For the fundamental matrix, we know that
This gives,
Consider the case of two cameras viewing an object such that the second camera differs from the first by a pure translation that is parallel to the
Since we have a pure translation parallel to the x axis, we get
Since we have a translation only about the x axis, our essential matrix becomes:
Our epipolar lines are given by
We get a normal vector of the epipolar line of the form
We know that the mapping from the 3D scene to the camera coordinates are dictated by the equation
For both cameras, we get:
Substituting equation 1 in equation 2 and eliminating
Hence, the fundamental matrix and essential matrix will be:
$$ E = [t_{rel}]{\times} R{rel}$$
$$ F = (K^{-1}) ^ T E K^{-1} = (K^{-1}) ^ T [t_{rel}]{\times} R{rel} K^{-1} $$
Suppose that a camera views an object and its reflection in a plane mirror. Show that this situation is equivalent to having two images of the object which are related by a skew-symmetric fundamental matrix. You may assume that the object is flat, meaning that all points on the object are of equal distance to the mirror (Hint: draw the relevant vectors to understand the relationship between the camera, the object, and its reflected image.)
Let us take the case where the image plane is perpendicular to the mirror. The camera views 2 copies of the object - 1 where the actual image is placed and 2) from the location of the reflected image. Mathematically, we can describe the image and its point on the camera plane as
Without loss of generality, let us assume that the camera plane is placed perpendicular to the mirror. This means that the reflection of the object will be a simple translation along the image plane. This means that there is no rotation involved between the two views from the discussion in the above paragraph.
We can hence write our essential matrix as $E = [t]{\times} R = [t]{\times}$ as
hence, our fundamental matrix then becomes
Let us now check if the fundamental matrix is skew symmetric (
Since
Hence, F is skew symmetric.
The recovered F is:
F = [[-0. 0. -0.2519]
[ 0. -0. 0.0026]
[ 0.2422 -0.0068 1. ]]
Error: 0.39895034989884903
Some outputs points from the display function.
The predicted F
F = [[ 0. 0. -0.201 ]
[ 0. -0. 0.0007]
[ 0.1922 -0.0042 1. ]]
Error = 0.5668901239522244
Let
Here it is clear that the points are detected correctly.
Correspondence matches. Computed using ORB descriptor and brute force matcher inbuilt in OpenCV.
Reconstructions:
In the above images, the rough shape of the spikes of the dinosaur can be seen. As the feature points are sparse, it is not very clearly visible, but the rough shape can be seen.
The original images are:
This is from the DinoSparseRing dataset from the provided links. images dinoSR0001 and dinoSR0002 are chosen. The relative change in angle between the 2 images is ~22.5 degrees. The fundamental matrix was estimated using RANSAC and 7 point algorithm.
(CG optimizer)
Errors: before 9170.318705933929, After 4145.34978423357
Errors: before 9170.318705933929, After 1828.3868371985636
Conceptual discussion was carried out with the following students:
- Siddharth Saha
- Ronit Hire