EigenHistograms: Using Low Dimensional Models of Color Distribution for Real Time Object Recognition Jordi Vitri´ a, Petia Radeva, and Xavier Binefa Centre de Visi´o per Computador, Dept. Inform` atica, Universitat Aut`onoma de Barcelona, 08193 Bellaterra (Barcelona),Spain, {jordi, petia}@cvc.uab.es Abstract. Distribution of object colors has been used in computer vi- sion for recognition and indexing. Most of the recent approaches to this problem have been focused on defining optimal spaces for representing pixel values that are related to physical models and that present some in- variance. We propose a new approach to identify individual object color distributions by using statistical learning techniques and to allow their compact representation in low dimensional spaces. This approach out- performs generic ”optimal” spaces when color illumination is constant, allowing changes in object pose and illumination direction. This approach has been tested for real time industrial inspection of multicolored objects. 1 Introduction Object appearance in an image is caused by many factors, including object pose, illumination directions and illumination color. Traditional recognition methods [7],[6] have mainly used geometric cues as shape models and feature matching for detecting and identifying objects in images, but they have not demonstrated the level of performance that allow them to be systematically used in real time for a large number of applications (i.e. face recognition). Recently, some new approaches based on the direct representation of object appearance have been developed for object recognition and pose estimation. Turk and Pentland [3] used principal component analysis to describe face patterns in a low dimensio- nal appearance space. Murase and Nayar in [4] have shown real time recognition of complex 3D objects based on Principal Component Analysis (PCA) of geo- metrical shape of the objects. The PCA approach is very appropriate for real time applications because of the low cost of the recognition algorithms, however it is limited to the analysis of the geometric shape and depends on the object’s pose. Color distributions can be efficiently used as signatures for object recogni- tion in the appearance-based framework. The earliest approach [2] showed the usefulness of color histograms for indexing large object databases independently of object’s pose. Most of the recent approaches focus on illumination color in- variance [1],[5] known as color constancy, but although these methods perform F. Solina and A. Leonardis (Eds.): CAIP ’99, LNCS 1689, pp. 17–24, 1999. c  Springer-Verlag Berlin Heidelberg 1999