4 New Developments and Open Questions
This paper has mainly concentrated on single view (object recognition) or two views
(motion based grouping, structure recovery) for a perspective (central) projection
camera. Similar uncalibrated models have been developed for a pushbroom camera
by Hartley and applied to SPOT images [11], and X-ray analysis.
The fundamental matrix is a bilinear relationship between two images that does
not depend on structure. In a similar manner there exist trilinear relationships [7]
between a triplet of images which are structure independent (i.e. only depending on
the camera internal parameters and relative dispositions). These relationships are
encapsulated in the “tri-focal tensor” which was originally developed by Shashua [25]
for point correspondences between three images, and shown by Hartley [12] to apply
also to line correspondences between image triplets. Recently, Carlsson [5] has
established a duality relation so that uncalibrated methods developed for m points
in n views can be applied directly to n + 4 points in m — 4 views. This has opened
up a host of possibilities for initialising structure and cameras from 4+ views.
There are a number of questions which the uncalibrated approach repeatably
generates, for example:
1. Is accuracy (e.g. of computed epipolar geometry) lost by taking an uncalibrated
route, as compared to the calibrated?
2. Where internal calibration is available or partially available, how can it be
used to advantage?
The answer to the first question appears to be no, the second is harder to answer so
succinctly, though calibration can be used to reduce search regions in uncalibrated
matching through image sequences. Other questions, such as how to carry out an
error analysis in projective 3-space, are less easy to handle.
Acknowledgements
This paper is based on the work of the Oxford Robotics Group, particularly Paul
Beardsley and Phil Torr. Financial support was provided by EU Esprit Project 6448
‘VIVA’.
References
[1] Armstrong, M., Zisserman, A. and Beardsley, P.A., Euclidean reconstruction from
uncalibrated images. Proc. British Machine Vision Conference, 1994. 2 3
[2] Ayache, N. and Faugeras, O.D. “HYPER: A New Approach for the Recognition and
Positioning of Two-Dimensional Objects,” IEEE Trans. PAMI, Vol. 8, No. 1, p.44-54,
January 1986.
[3] Beardsley, P.A., Zisserman, A. and Murray, D.W., Navigation using affine structure
and motion. Proc. ECCV, LNCS 800-801, Springer-Verlag, 1994.
Zisserman - 13