Exploiting object characteristics using custom features for boosting-based classification

Abstract

Typical feature pools used to train boosted object detectors contain various redundant and unspecific information which often yield less discriminative detectors. In this paper we introduce a feature mining algorithm taking domain specific knowledge into account. Our proposed feature pool contains rectangular shaped features generated from an image clustering algorithm applied on the mean image of the object training set. A combination of two such spatially separated rectangular regions yields a set of features which have a similar evaluation time like classical Haar-like features, but are much smarter (automatically) selected and more discriminative since image correlations can be more consequently exploited. Overall, training is faster and results in more selective detectors showing improved precision. Several experiments demonstrate the gain when using our proposed feature set in contrast to standard features. © 2013 Springer-Verlag.