Abstract. We present a method to learn object class models from unlabeled and unsegmented cluttered scenes for the purpose of visual object recognition. We focus on a particular type of model where objects are represented as flexible constellations of rigid parts (features). The variability within a class is represented by a joint probability density function (pdf) on the shape of the constellation and the output of part detectors. In a first stage, the method automatically identifies distinctive parts in the training set by applying a clustering algorithm to patterns selected by an interest operator. It then learns the statistical shape model using expectation maximization. The method achieves very good classification results on human faces and rear views of cars. 1 Introduction and Related Work We are interested in the problem of recognizing members of object classes, where we define an object class as a collection of objects which share characteristic features or parts that are visually similar and occur in similar spatial configurations. When building models for object classes of this type, one is faced with three problems (see Fig. 1).

Abstract. We propose a novel framework for detecting human faces based on correspondences between triplets of detected local features and their counterparts in an affine invariant face appearance model. The method is robust to partial occlusion, feature detector failure and copes well with cluttered background. Both the appearance and configuration probabilities are learned from examples. The method was tested on the XM2VTS database and a limited number of images with cluttered background with promising results – 2 % false negative rate – was obtained. 1

We propose a robust method for face detection based on the assumption that face can be represented by arrangements of automatically detectable discriminative regions. The appearance of face is modeled statistically in terms of local photometric information and the spatial relationship of the discriminative regions. The spatial relationship between these regions serves mainly as a preliminary evidence for the hypothesis that a face is present in a particular position. The nal decision is carried out using the complete information from the whole image patch. The results are very promising.

The aim of this thesis proposal is to explore and develop a system that robustly solves the problem of human face localization with minimal amount of constraints on scene conditions (illumination, head pose etc.) and facial expressions. The appearance of the face is modeled locally in terms of discriminative regions and their spatial relationship. Most of the free parameters are learned from training examples and as little as possible information from a human expert is needed.

In this paper, we propose a real-time face and eye detection algorithm for video surveillance and human computer interface. Different types of facelnon-face classifiers are hierarchically combined for reliability and real-time performance. Each classifier is connected by consisdering the accuracy and translationlscale sensitivity of the previous classifier. First, face candidates are extracted using similarity matching of Gabor filter responses in M-style grid. Then, a hierarchy of SVM classifiers trained on PCA subspaces is applied to the candidates. Two steps of SVMs are trained on different number of PCA features and resolution images. The combination of classifiers based on different types of features, frequency and intensity, significantly reduces the false positives due to complementary characteristics of their domains in classification. Coarse-to-fine search by three steps of detection also reduces run-time complexity. Our system can speed up conventional SVM classifier by a factor of 40 resulting in comparable face detection rate (FDR). In addition, the center positions of both eyes are efficiently detected by iterative binary thresholding method with contour tracing in the accurately localized face region. Experimental results on the test set taken under office illumination show the accuracy with FDR of 98%, 0.5 % false positives, and eye detection rate (EDR) of 99%. Our current system detects a face in runtime of 250ms. 1