This paper presents an empirical analysis of extensions to the Viola and Jones object detection framework, focusing on two main improvements: the introduction of rotated Haar-like features and a detailed comparison of boosting algorithms. The rotated features, which include 45° rotated rectangles, enhance the system's ability to detect objects with diagonal structures, leading to a 10% reduction in false alarm rates. The analysis of boosting algorithms (Discrete, Real, and Gentle Adaboost) shows that Gentle Adaboost, using small CART trees as weak classifiers, outperforms the others in terms of detection accuracy and computational complexity. The paper also discusses the optimal input pattern size for frontal face detection, which is 20x20, and the benefits of using more complex weak classifiers like small CART trees over stumps. The complete system and a trained face detector are available in the Open Computer Vision Library.This paper presents an empirical analysis of extensions to the Viola and Jones object detection framework, focusing on two main improvements: the introduction of rotated Haar-like features and a detailed comparison of boosting algorithms. The rotated features, which include 45° rotated rectangles, enhance the system's ability to detect objects with diagonal structures, leading to a 10% reduction in false alarm rates. The analysis of boosting algorithms (Discrete, Real, and Gentle Adaboost) shows that Gentle Adaboost, using small CART trees as weak classifiers, outperforms the others in terms of detection accuracy and computational complexity. The paper also discusses the optimal input pattern size for frontal face detection, which is 20x20, and the benefits of using more complex weak classifiers like small CART trees over stumps. The complete system and a trained face detector are available in the Open Computer Vision Library.