Efficient multiclass object detection by a hierarchy of classifiers

We consider detecting object instances from multiple classes on grayscale images. Traditional approaches learn a classifier for each class separately and apply each of them in an exhaustive search over positions and scales. We achieve an efficient detection by organizing the search coarse-to-fine based on a hierarchical partitioning of the entire hypothesis space, the set of all possible object instances, so that groups of hypotheses can be pruned simultaneously without evaluating each one individually. In this paper, we develop an algorithm to jointly learn the hierarchy along with a classifier at each node by exploring the common parts shared among a group of object instances at all levels in the hierarchy. We also show how the confusions of the initial coarse-to-fine search can be resolved by comparing pairs of conflicting detections using cheap binary classifiers. The whole process is illustrated by detecting and recognizing handwritten digits.