In this paper, we describe a new feature called Separate Haar (Sep-Haar) feature for fast and accurate face detection. There are three key contributions. "Separate Haar feature" adds a negligible area for the rectangular Haar feature window, by which we can improve the feature extraction efficiency; the corresponding algorithm for selecting the best width of such negligible area is realized by reducing the total number of learned features to reduce the memory used; and experiment result shows that the proposed Sep-Haar feature can achieve best false alarm rate using less number of features in Adaboost algorithm compared with traditional Haar feature. Based on the result, we propose a new classifier that, by using the proposed Sep-Haar features, it can give smaller false alarm rate at each stage, use less number of stages, and at the same time give improved hit rate within the same detection time consummed.