Recently we have proposed a novel mathematical model for cascade structure and developed a new boosting method called FCBoost for training a cascade of classifiers. FCBoost designs a cascade based on its over all performance and adjust all cascade settings such as number of stages, number of weak learners in each stage and stage thresholds automatically.
Also we have developed a new cost sensitive Boosting method called Log-CS-Boost whose input is the desired detection rate rather than non-intuitive cost sensitive factors. Using this method there is no need for try and error for adjusting those factors.
Using these methods we have developed a cascade based on FCBoost and a cascade with combination of Log-CS-Boost and embedded cascade structure.Blew are some demos of the classifiers on MIT-CMU face detection data set and YouTube Celebrity data set set for face tracking.