Standard cameras are central—all rays pass through a single optical center. However, multi-camera rigs (like those on self-driving cars) are non-central. Rays from different cameras do not share a common convergence point.
In a Generalized Camera Model (GCM), we represent each light ray as a Plücker line $\mathbf{L} = (\mathbf{u}, \mathbf{q})$.
Where $\mathbf{c}$ is the origin of the specific camera in the rig coordinate system.
Because there is no single center of projection, a simple $3 \times 3$ homography cannot relate two views of a scene if the rig moves. The transformation depends on depth, creating a parallax effect that traditional central models cannot resolve.