Boosting 2-Thresholded Weak Classifiers over Scattered Rectangle Features for Object Detection

Abstract

In this paper, we extend Viola and Jones’ detection framework in two aspects. Firstly, by removing the restriction of the geometry adjacency rule over Haarlike feature, we get a richer representation called scattered rectangle feature, which explores much more orientations other than horizontal, vertical and diagonal, as well as misaligned, detached and non-rectangle shape information that is unreachable to Haar-like feature. Secondly, we strengthen the discriminating power of the weak classifiers by expanding them into 2-thresholded ones, which guarantees a better classification with smaller error, by the simple motivation that the bound on the accuracy of the final hypothesis improves when any of the weak hypotheses is improved. An optimal linear online algorithm is also proposed to determine the two thresholds. The comparison experiments on MIT+CMU upright face test set under an objective detection criterion show that the extended method outperforms the original one.