Hybrid Euclidean-and-Riemannian Metric Learning for Image Set Classification

Face recognition on large-scale video in the wild with hybrid Euclidean-and-Riemannian metric learning

“…In fact, this work is an extension of our previous work [33]. The differences between this work and the conference paper are as follows: (1) this paper extends the Single Gaussian Model (SGM) in the conference version to Gaussian Mixture Model (GMM), which is essentially a general version of SGM, for modeling the Gaussian distribution.…”

Report on the FG 2015 Video Person Recognition Evaluation

“…to HERML-DeLF), which is basically the HERML method [13] for image set classification with image features learned by a deep neural network 4 For the feature learning part of our HERML-DeLF method, a deep convolutional neural network (DCNN) model is trained on 256 by 256 pixel face images. For a fair comparison, we normalize the face images using eye positions provided by the organizers of PaSC [2].…”

“…Using the DCNN features, the HERML method [13] is then used to compute video similarity by fusing three different set-based video representations. Specifically, for each video, the DCNN features of all video frames are first pooled respectively by sample mean, sample covariance matrix and Gaussian model, which form three types of setbased video representations.…”

“…Specifically, for each video, the DCNN features of all video frames are first pooled respectively by sample mean, sample covariance matrix and Gaussian model, which form three types of setbased video representations. Then, by applying the kernel functions proposed in [13] for set-based representations, three kernel matrices are computed and fed separately into kernel linear discriminant analysis (KLDA) [1]. Here, instead of the original metric fusing method in [13], we exploit the KLDA to learn three projective functions respectively [28].…”

“…Then, by applying the kernel functions proposed in [13] for set-based representations, three kernel matrices are computed and fed separately into kernel linear discriminant analysis (KLDA) [1]. Here, instead of the original metric fusing method in [13], we exploit the KLDA to learn three projective functions respectively [28]. The resulting projective functions are then used to produce three 440 dimensional feature vectors for each video.…”