This paper presents an empirical analysis of detection cascades of boosted classifiers for rapid object detection. The authors introduce and analyze two extensions to the original approach proposed by Viola et al. First, they introduce a novel set of rotated haar-like features, which significantly enrich the simple features and can be calculated efficiently. These features reduce the false alarm rate by about 10% at a given hit rate. Second, they analyze different boosting algorithms (Discrete, Real, and Gentle Adaboost) and weak classifiers, finding that Gentle Adaboost with small CART trees outperforms the other algorithms in terms of detection accuracy and computational complexity. The feature set is based on the over-complete haar-like features from previous work and includes edge, line, and center-surround features. These features are scaled independently in vertical and horizontal directions to generate a rich set of features. The authors also introduce a fast feature computation method using auxiliary images, which allows for efficient calculation of feature values. The paper also discusses the use of boosting as a basic classifier, which combines the performance of many weak classifiers to produce a powerful 'committee'. The authors compare three boosting algorithms and find that Gentle Adaboost performs best. They also compare different input pattern sizes and find that 20x20 is optimal for frontal face detection. The authors also compare different feature sets and find that the extended haar-like features improve detection performance. They also discuss the use of different weak classifiers, finding that small CART trees outperform stumps. The paper concludes that Gentle Adaboost outperforms Discrete and Real Adaboost in terms of detection accuracy and computational complexity. The authors also introduce an extended set of haar-like features, which improve detection performance, especially for objects with diagonal structures. The complete training and detection system is available in the Open Computer Vision Library.This paper presents an empirical analysis of detection cascades of boosted classifiers for rapid object detection. The authors introduce and analyze two extensions to the original approach proposed by Viola et al. First, they introduce a novel set of rotated haar-like features, which significantly enrich the simple features and can be calculated efficiently. These features reduce the false alarm rate by about 10% at a given hit rate. Second, they analyze different boosting algorithms (Discrete, Real, and Gentle Adaboost) and weak classifiers, finding that Gentle Adaboost with small CART trees outperforms the other algorithms in terms of detection accuracy and computational complexity. The feature set is based on the over-complete haar-like features from previous work and includes edge, line, and center-surround features. These features are scaled independently in vertical and horizontal directions to generate a rich set of features. The authors also introduce a fast feature computation method using auxiliary images, which allows for efficient calculation of feature values. The paper also discusses the use of boosting as a basic classifier, which combines the performance of many weak classifiers to produce a powerful 'committee'. The authors compare three boosting algorithms and find that Gentle Adaboost performs best. They also compare different input pattern sizes and find that 20x20 is optimal for frontal face detection. The authors also compare different feature sets and find that the extended haar-like features improve detection performance. They also discuss the use of different weak classifiers, finding that small CART trees outperform stumps. The paper concludes that Gentle Adaboost outperforms Discrete and Real Adaboost in terms of detection accuracy and computational complexity. The authors also introduce an extended set of haar-like features, which improve detection performance, especially for objects with diagonal structures. The complete training and detection system is available in the Open Computer Vision Library.