This paper studies the usefulness of the projective approach to structure from
motion (SFM).  We conduct an experimental and algorithm---independent  comparison
of projective versus Euclidean reconstruction. Our results show that
Euclidean reconstruction is essentially as accurate as projective
reconstruction, even for the pure projective structure and significant
calibration errors.  Thus calibration error is not a compelling
motivation for the projective framework. But projective optimization
is more reliable than its Euclidean equivalent: we find that the
Levenberg--Marquardt algorithm has less of a local--minimum problem in
the projective case.  We describe several techniques that enhance the
convergence of the Levenberg--Marquardt algorithm. Most importantly, we find that it
is crucial to exploit the compactness of projective space to achieve
fast, reliable convergence of SFM optimization algorithms---even in
the pure Euclidean framework.  We also analyze the problem of
determining a stable basis set of 5 points for projective reconstruction.