We use a multi-color classification method introduced by Wolf et al. (2001)
to reliably identify stars, galaxies and quasars in the up to 16-dimension
al color space: provided by the filter set of the Calar Alto Deep Imaging S
urvey (CADIS). The samples of stars, galaxies and quasars obtained this way
have been used for dedicated studies which are published in separate paper
s. The classification is good enough to detect quasars rather completely an
d efficiently without confirmative spectroscopy. The multi-color redshifts
are accurate enough for most statistical applications, e.g. evolutionary st
udies of the galaxy luminosity function. Also, the separation between stars
and galaxies reaches deeper than with morphological criteria, so that stud
ies of the stellar population can be extended to fainter levels. We charact
erize the dataset presently available on the CADIS 1 h-, 9 h- and 16 h-fiel
ds. Using Monte-Carlo simulations we model the classification performance e
xpected for CADIS. We present a summary of the classification results on th
e CADIS database and discuss unclassified objects. More than 99% of the who
le catalog sample at R < 22 (more than 95% at R < 23) are successfully clas
sified matching the expectations derived from the simulations. A small numb
er of peculiar objects challenging the classification is discussed in detai
l. Spectroscopic observations are used to check the reliability of the mult
i-color classification (6 mistakes among 151 objects with R < 24). From the
se, we also determine the accuracy of the multi-color redshifts which are r
ather good for galaxies (<sigma>(z) approximate to 0.03) and useful for qua
sars. We find that the classification performance derived from the simulati
ons compares well with results from the real survey. Finally, we locate are
as for potential improvement of the classification.