Face verification is the task of determining whether two given face images represent the same person or not. It is a very challenging task, as the face images, captured in the uncontrolled environments, may have large variations in illumination, expression, pose, background, etc. The crucial problem is how to compute the similarity of two face images. Metric learning has provided a viable solution to this problem. Until now, many metric learning algorithms have been proposed, but they are usually limited to learning a linear transformation. In this paper, we propose a nonlinear metric learning method, which learns an explicit mapping from the original space to an optimal subspace using deep Independent Subspace Analysis (ISA) network. Compared to the linear or kernel based metric learning methods, the proposed deep ISA network is a deep and local learning architecture, and therefore exhibits more powerful ability to learn the nature of highly variable dataset. We evaluate our method on the Labeled Faces in the Wild dataset, and results show superior performance over some state-of-the-art methods.

In Still-to-Video (S2V) face recognition, only a few high resolution images are registered for each subject, while the probe is video clips of complex variations. As faces present distinct characteristics under different scenarios, recognition in the original space is obviously inefficient. Thus, in this paper, we propose a novel discriminant analysis method to learn separate mappings for different scenario patterns (still, video), and further pursue a common discriminant space based on these mappings. Concretely, by modeling each video as a manifold and each image as point data, we form the scenario-oriented mapping learning as a Point-Manifold Discriminant Analysis (PMDA) framework. The learning objective is formulated by incorporating the intra-class compactness and inter-class separability for good discrimination. Experiments on the COX-S2V dataset demonstrate the effectiveness of the proposed method.

Recognizing human action in complex scenes is a challenging problem in computer vision. Some action-unrelated concepts, such as camera position features, could significantly affect the appearance of local spatio-temporal features, and therefore the performance of low-level features based methods degrades. In this letter, we define the action-unrelated concept: the position of camera as high-level features. We observe that they can serve as a prior to local spatio-temporal features for human action recognition. We encode this prior by modeling interactions between spatio-temporal features and camera position features. We infer camera position features from local spatio-temporal features via these interactions. The parameters of this model are estimated by a new max-margin algorithm. We evaluate the proposed method on KTH, IXMAS and Youtube actions datasets. Experimental results show the effectiveness of the proposed method.