Abstract. This paper shows an approach for automatic learning of efficient representations for robust image features. A video sequence of a 3D scene is processed using structure-from-motion algorithms, which provides a long validated track of robust 2D features for each tracked scene region. Thus each tracked scene region defines a class of similar feature vectors forming a volume in feature space. The variance within each class results from different viewing conditions, e.g. perspective, lighting conditions, against which the feature is not invariant. We show on synthetic and on real data that making use of this class information in subspace methods, a much sparser representation can be used. Furthermore, less computational effort is needed and more correct correspondences can be retrieved for efficient computation of the pose of an unknown camera image than in previous methods.