Zisserman - 1
Uncalibrated Vision
Andrew Zisserman
Robotics Research Group
Department of Engineering Science
University of Oxford
Oxford 0X1 3PJ, UK.
1 Introduction
Uncalibrated vision refers to a body of techniques whose starting point is that a
camera is a projective sensor, modelled as a linear map between 3-space and the
image. In particular the camera calibration (the interior orientation, or intrinsic
parameters) is unknown. Subsequent computations are based solely on projective
geometry (intersection of lines with planes, for example) as opposed to Euclidean
geometry where angles can be measured.
In contrast, the starting point for conventional photogrammetry (and much of
computer vision) is to precisely model and determine the interior orientation of the
camera. The camera is then modelled as a Euclidean sensor, able to measure the
angle between rays.
The uncalibrated approach is a necessity in many computer vision applications.
There are often cases where calibration information is not available (at least ini
tially), for example video sequences or archive photographs, or cases where an initial
calibration will be lost, for example a camera mounted (and shaken) on a moving
robot arm or AGV, or deliberate change of focal length (zooming). It might be pos
sible to guess values for the internal parameters, or parameterise their change (under
zooming). However, as will be demonstrated in the sequel, calibration is simply not
necessary for many important applications and tasks. Furthermore, where projec
tive geometry is insufficient for a task, uncalibrated vision makes explicit the precise
geometric strata (affine, metric) required. The approach is in concord with the
‘Active Vision’ paradigm [4] of computer vision, where only sufficient information
necessary for the task in hand is extracted from the image or image sequence.
Where affine or metric information is required, the flavour of the computation
again differs from that of the photogrammetrist. A traditional approach would
solve directly for calibration (using the DLT for example) or employ particular scene
knowledge (such as the Euclidean coordinates of a number of points, or that certain
lines are parallel). In contrast the uncalibrated approach harnesses additional knowl
edge of the motion (it may be pure translational for example) or camera to generate