Tomokazu Kawahara

Recognizing Faces of Moving People by Hierarchical Image-Set Matching

This paper proposes a novel method for recognizing faces in a cluster of moving people. In this task, there are two problems caused by motion, which are occlusions, and changes in facial pose and illumination. Multiple cameras are used to acquire near-frontal faces to avoid occlusions and profile faces. The hierarchical image-set matching (HISM) creates a distribution for each individual by integrating a set of face images of the same individual acquired from the multiple cameras. By adopting a method for comparing between test and training distributions in identification, variation in pose and illumination is alleviated, and good recognition accuracy can be obtained. Experimental results using video sequences containing 349 people show that the proposed method achieves high recognition performance compared with conventional methods, which use frame-by-frame identification and a distribution obtained from a single camera.

Automatic ship recognition robust against aspect angle changes and occlusions

We propose a novel automatic recognition method of ships in images produced by inverse synthetic aperture radar (ISAR). It has robustness against ship deformation due to changing aspect angles and loss of ship parts caused by occlusion. To deal with the deformation and the loss, we extract a feature vector from a ship in an ISAR image by Co-occurrence Histograms of Oriented Gradients (CoHOG). An ISAR ship image is divided into multiple blocks and CoHOG is extracted from pairs of quantized gradient orientations in each of the blocks. Quantized orientations are not changed by slight deformation of a ship. This property derives that CoHOG has robustness against the deformation due to changing aspect angles. Variation of CoHOG caused by occlusion is small since gradient orientations are changed in only blocks including occluded parts. On the other hand, combinations of pairs of orientations make dimension of CoHOG extremely high. We calculate a similarity between two ships using Random Ensemble Metric (REMetric), which is a metric learning method for a high dimensional feature space. It has multiple Support Vector Machines (SVM) learnt from randomly subsampled training data, and calculates a similarity between two vectors from results of SVM of these two vectors. Since SVM finds a hyperplane which has maximum margin between two classes, its classification performance is high even through dimension of a feature space is high. Through experiments with simulated ISAR ships images, we show our method has robustness against aspect angle changes and occlusion, and it has higher performance than a conventional method.

On the behavior of kernel mutual subspace method

Optimizing the parameters of kernel methods is an unsolved problem. We report an experimental evaluation and a consideration of the parameter dependences of kernel mutual subspace method (KMS). The following KMS parameters are considered: Gaussian kernel parameters, the dimensionalities of dictionary and input subspaces, and the number of canonical angles. We evaluate the recognition accuracies of KMS through experiments performed using the ETH- 80 animal database. By searching exhaustively for optimal parameters, we obtain 100% recognition accuracy, and some experimental results suggest relationships between the dimensionality of subspaces and the degrees of freedom for the motion of objects. Such results imply that KMS achieves a high recognition rate for object recognition with optimized parameters.

Feature Vector Compression Based on Least Error Quantization

We propose a distinctive feature vector compression method based on least error quantization. This method can be applied to several biometrics methods using feature vectors, and allows us to significantly reduce the memory size of feature vectors without degrading the recognition performance. In this paper, we prove that minimizing quantization error between the compressed and original vectors is most effective to control the performance in face recognition. A conventional method uses non-uniform quantizer which minimizes the quantization error in terms of L2-distance. However, face recognition methods often use metrics other than L2-distance. Our method can calculate the quantized vectors in arbitrary metrics such as Lp-distance (0