You are on page 1of 2

Camera matrix (3x4) is an element of projective space

A consequence of this relation is that also C can be seen as an element of a pro


jective space; two camera matrices are equivalent if they are equal up to a scal
ar multiplication.

3 reasons for homogenous coordinates:


1. Projective space: all equivalent up to a scale factor, since a given point co
uld be at any depth along the ray from the point to the camera center
2. Allows you to use linear algebra for non-linear stuff: in linear algebra, you
can't represent dividing by Z. But in homogenous coordinates, you can multiply/
divide by any nonzero Z and the result is the same
3. You can represent points at infinity by (x,y,z=0), since dividing all coordin
ates by last coordinate (z) tells you they are at infinity.

One method of how to determine Camera matrix, P (3x4):


PC = 0, i.e. the camera matrix acting on the camera center maps to 0 (the null s
pace). So use SVD: C is the vector corresponding to the smallest singular value.
Next, use the fact that the 3x3 internal calibtration matrix, M, is non-singular
and is the upper triangular matrix K multiplied by a rotation matrix R, which i
s an orthogonal matrix since it is a rotation matrix.
So use RQ factorizxtion (like QR factorization but reversed) to separate out R a
nd K.

intrinsic camera transformation occurs post-projection. One notable result of th


is is that intrinsic parameters cannot affect visibility
occluded objects cannot b
e revealed by simple 2D transformations in image space.

Calculating Fundamental Matrix:


use x'^T F x = 0
split into 9 equations
but the solution is homogenous (0 vector on right), so one of the degrees of fre
edom is arbitrary: all results within a scale factor are the same
So choose one of the paramters and arbitrarily set = 1
So do normalized 8 point algorithm

things that CS131 says are in CS231:

graph cut segmentation


spectral clustering

You might also like