A biased selection strategy for information recycling in Boosting cascade visual-object detectors

We study the problem of information recycling in Boosting cascade visual-object detectors. It is believed that information obtained in the earlier stages of the cascade detector is also beneficial for the later stages, and that a more efficient detector can be constructed by recycling the existing information. In this work, we propose a biased selection strategy that promotes re-using existing information when selecting weak classifiers or features in each Boosting iteration. The strategy used can be interpreted as introducing a cardinality-based cost term to the Boosting loss function, and we solve the learning problem in a step-wise manner, similar to the gradient-Boosting scheme. Our work provides an alternative to the popular sparsity-inducing norms in solving such problems. Experimental results show that our method is superior to the existing methods.