Fast Frontal-View Face Detection Using a Multi-path Decision Tree

Abstract

This paper presents a novel approach to the real-time detection of frontal faces in static grey images. The faces may be rotated in-plane from -60° to 60°which suffices for most practical applications. The detector consists of a multi-path decision tree, with two different types of node classifiers. Each tree node has the capability to reject the current subwindow or to classify it in one out of 3 rotation classes. The node classifiers are designed so that most of the non-face locations are eliminated at early nodes, leaving the harder cases to later nodes. The improvement at each node is twofold, higher angular resolution and successive rejection of non-face locations. The early nodes consist of sparse linear feature nets defined in the gradient direction domain which are very efficient to compute. So many of the non-face locations are rejected with low computational effort. The classifiers in the last node apply linear feature nets using spatially sampled grey value features. The proposed detector design allows to focus onto interesting image regions very rapidly allowing for a real-time system on standard PCs which can process approximately 25 frames per second. We present experimental results of the system on the widely used CMU Database (frontal and rotated set) and the BioID Database where an overall detection rate of 91% is achieved.