A new pattern recognition approach to face recognition is presented that can deal with drastic differences in the appearance of a face. Given a pair of sample sets of facial images with potential correspondences, each being drawn from a distinctive distribution, the algorithm reliability finds correspondences over those different distributions. Unlike the traditional approaches that model the face images as having a consistent distribution and so use the same feature extraction function for both of the image sets, the new method applies to each sample a function specific to the distribution from which it is drawn. This function is derived by maximizing a newly defined class-separability criterion over the different distributions. Results of face recognition are presented on images including drivers' license pictures. Drastic improvements are shown over algorithms based on the traditional Fisher's discriminant analysis.

This paper describes a new method for face recognition under drastic changes of the imaging processes through which the facial images are acquired. Unlike the conventional methods that use only the face features, the present method exploits the statistical information of the variations between the face image sets being compared, in addition to the features of the faces themselves. To incorporate both of the face and perturbation features for recognition, we develop a technique called weak orthogonalization. Of the two subspaces that transforms the given two overlapped subspaces so that the volume of the intersection of the resulting two subspaces is minimized. Matching operations are performed in the transformed face space that has thus been weakly orthogonalized against perturbation space. Experimental results on real pictures of the frontal faces from drivers' licenses show that the new algorithm improves the recognition performance over the conventional methods. We also demonstrate the effectiveness of our method on image sets with changes in viewing geometry.

Novel algorithms for directly checking the planarity and the rigidity of object surfaces are described. Unlike the conventional methods, our algorithms need not recover the parameters included in the equations constraining the motion of the planar or the rigid surfaces. Instead, given the correspondences of the 2D image feature points over a small number of 2D images. They directly evaluate those properties by analyzing the shapes of the feature distributions in the feature space constructed by combining the 2D features. Then, they measure how well their shapes fit the types of those for planar or rigid surfaces that are underlyingly characterized by the conventional constraint equations. Experimental results on both simulated and real data demonstrate the effectiveness of the proposed algorithms.

We describe an efficient algorithm for recognizing 3D objects by combining photometric, and geometric invariants. A photometric property is derived, that is invariant to the changes of illumination and to relative object motion with respect to the camera and/or the lighting source in 3D space. We argue that conventional color constancy algorithms can not be used in the recognition of 3D objects. Further we show recognition does not require a full constancy of colors, rather, it only needs something that remains unchanged under the varying light conditions and poses of the objects. Combining the derived color invariant and the spatial constraints on the object surfaces, we identify corresponding positions in the model and the data space coordinates, using centroid invariance of corresponding groups of feature positions. Tests are given to show the stability and efficiency of our approach to 3D object recognition.>

Novel algorithms for object recognition are described that directly recover the transformations relating the image to its model. Unlike methods fitting the conventional framework, these new methods do not require exhaustive search for each feature correspondence in order to solve for the transformation. Yet they allow simultaneous object identification and recovery of the transformation. Given hypothesized corresponding regions in the model and data (2D views)-which are from planar surfaces of the 3D objects-these methods allow direct computation of the parameters of the transformation by which the data may be generated from the model. We propose two algorithms: one based on invariants derived from no higher than second order moments of the image, the other via a combination of the affine properties of geometrical and differential attributes of the image. Empirical results on natural images demonstrate the effectiveness of the proposed algorithms. We demonstrate in particular that the differential method is quite stable against perturbations-although not without some error-when compared with conventional methods.

This paper presents an efficient and robust alignment algorithm for recognizing 3D objects. We first show that for features from planar surfaces which undergo linear transformations in space, there exists a class of transformations that yield projections invariant to the surface motions, up to rotations in the image field. To use this property in recognition, we propose a new alignment approach based on centroid alignment of corresponding feature groups built on these invariant projections of planar surfaces. This method uses only a single pair of 2D model and data pictures for recognizing a 3D object. Some results on natural pictures are given.