Camera Calibration for 3D Reconstruction and View Transformation - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 110

Performance analysis and evaluation
In Kumar & Hanson (1989), a mathematical analysis and experiments were
carried out to develop a closed-form function to express the uncertainty of the
calibration process. It was shown theoretically and experimentally in Lai (1993)
that the offset of the image center does not significantly affect the determination
of the position and orientation of a coordinate frame. An experimental perfor-
mance analysis of a traditional calibration algorithm was conducted in Scott &
Mohan (1995), where the result is evaluated with a geometric interpretation.
Recently, the influence of noisy measurements on the camera calibration matrix
~ and on all linear camera calibration parameters in p l has been analyzed
theoretically and verified using Monte Carlo simulations (Kopparapu & Corke,
2001).
Accuracy evaluation is a crucial part in the development of new camera
calibration algorithms. The 3-D reconstruction error and the image-plane
projection error are probably the most popular evaluation criteria employed.
However, due to the differences in image digitization and vision set-up, some
normalization should be performed on the criteria to get comparable results for
different calibration systems (Hartley & Zisserman, 2000).
Self-Calibration
Self-calibration is the process of determining camera parameters directly from
uncalibrated video sequences containing a sufficient number of frames. These
video sequences are generated from a single camera wandering in a still 3-D
scene, or from multiple cameras at different poses imaging a still 3-D scene, or
a single camera observing a moving object. All of these situations are equivalent
to the case of “multiple cameras at different poses imaging a still 3-D scene.”
If an arbitrary projective transformation is applied on all input data sets, the
relative relations among them will not be altered. Thus, with sufficient point
correspondence relations among available views, the camera parameters can
only be determined up to a projective ambiguity (Hartley & Zisserman, 2000).
However, in order to obtain the camera geometry, only a Euclidean ambiguity is
allowed. This is why in the self-calibration process certain constraint or a priori
information has to be assumed to upgrade the projective reconstruction to a
Euclidean one (Faugeras, 1994). Generally speaking, the following three types
of constraints or information can be employed: