Boosting algorithms for detector cascade learning

The problem of learning classifier cascades is considered. A new cascade boosting algorithm, fast cascade boosting (FCBoost), is proposed. FCBoost is shown to have a number of interesting properties, namely that it 1) minimizes a Lagrangian risk that jointly accounts for classification accuracy and speed, 2) generalizes adaboost, 3) can be made cost-sensitive to support the design of high detection rate cascades, and 4) is compatible with many predictor structures suitable for sequential decision making. It is shown that a rich family of such structures can be derived recursively from cascade predictors of two stages, denoted cascade generators. Generators are then proposed for two new cascade families, last-stage and multiplicative cascades, that generalize the two most popular cascade architectures in the literature. The concept of neutral predictors is finally introduced, enabling FCBoost to automatically determine the cascade configuration, i.e., number of stages and number of weak learners per stage, for the learned cascades. Experiments on face and pedestrian detection show that the resulting cascades outperform current state-of-the-art methods in both detection accuracy and speed.