A combination of generative and discriminative approaches to object detection

This paper presents a new simple algorithm which combines generative and discriminative approaches to object detection. The research makes two key contributions. The first contribution is the introduction of a new algorithm called the DT (Decomposition-Tree) which is capable of clustering on the manifold of object patterns(using Gaussian clusters) and determining the thresholds of each cluster by using hard samples which are selected during learning. The second contribution is that the learning time of the DT algorithm has been reduced rapidly. Because the DT algorithm shows spatial relationships of training patterns in the form of a tree, it requires relearning rather than new learning. To evaluate the performance of the proposed object detection algorithm, we experimented with face detection. The DT algorithm yields face detection performance comparable to that of the best previous systems [4]