The objective of this work is to detect a wide variety of object
classes (such as birds and trains) in an image.  We propose a novel
object class model based on multiple kernel combinations, with each
kernel representing a different feature and spatial
information. Objects are detected using a sliding window.  We make
three contributions: (i) we improve on the learning method of Varma &
Ray [ICCV 2007] by partitioning the search space, and using Varma &
Ray's algorithm as an initialization for a local optimization over a
validation set; (ii) we enable efficient learning of additional
(synthesized) training data by using bootstrapping; and (iii) we
introduce a hierarchical model based on class confusions on the
validation set.

We develop detectors for all the animal and vehicle categories (12
in total) of the PASCAL VOC07 detection challenge.  On the test data
we surpass the state-of-the-art by at least 50\% on all categories,
and by more than 100\% on some of them (e.g. for birds the AP
increases from $0.098$ to $0.638$).\vspace{-0.3cm}