On features ordering for rapid object detection

Abstract

Consider a real time classification problem which requires testing the value of a large number of features. Let Tj be the average computation time of the jth feature, and let Rj be its average rejection probability. In many real-world problems, such as defect detection on printed circuits or plasma screens, minimizing classifier execution time is of paramount importance because of its effect on the entire productions and QA process. We show that if features are independent, the best way to arrange them is according to the ratio Rj=Tj. This yields a minimal-time algorithm for features ordering in object detection framework when feature values are independent of each other, and motivates a simple algorithm for ordering features that are not independent. The algorithm obtains comparable classification results to [15], but is twice as fast.