Out-of-Plane Rotated Object Detection using Patch Feature based Classifier

In this paper, we presented a method to extend the weak classifiers that we have previously developed called the square patch feature for out-of-plane rotated object detection. The square patch feature is as discriminative as the popular Viola-Jones Haar-like classifier and is faster. Out-of-plane detection without any extra sample data is possible due to the point based representation of the patch feature. The feature points in the classifier data trained from a frontal face can be rotated by assuming that they are mapped on a surface of the object of interest. For simplification object of interest such as the face can be assumed to be flat. The method was tested for face detection problem. A face detector was trained using 4916 face images and rotated for out-of-plane detection. The detection rate of the out-of-plane detection is 71% with a false positive of 189.