Though perspective projections give accurate models for a wide range of existing cameras, the mapping from the object point to the image point is non-linear due to the scaling factor . In order to make the projection model more mathematically tractable, affine cameras have been introduced.
Now we pick up one arbitrary point in 3D space and let be the image of observed by i-th camera. We call a reference point. We also introduce a 23 matrix and a 3-vector which are composed from first two rows and third row of respectively. Then the depths of and for i-th camera are given by and respectively with . Using these entities, we approximate the perspective projection (1) as
where we assumed that and . Translating both image and object coordinates by and , we have an affine camera model:
which is a first-order approximation obtained from Taylor expansion of the perspective camera model around the reference point . The reference point may be chosen from the object points or, in many cases, be set to their centroid, i.e. . In the latter, the image of in i-th camera coincides, to a first-order approximation, with the centroid of the images of the object points, that is, .
If the intrinsic camera parameters in are known, we can set by using appropriate canonical image coordinates. Under this condition, three special types listed below are important instances of the affine camera;
which is refered to as a paraperspectve camera model.
where the distances of the reference point from the cameras are normalized to unity.