Ilias Theodorakopoulos

Pose-based human action recognition via sparse representation in dissimilarity space

A framework for pose-based human action recognition is proposed, where coordinates of joints are considered as inputs.By comparing corresponding trajectories, actions are represented by vectors of dissimilarities to a set of prototypes.The task of recognition is performed into the dissimilarity space.The newly created UPCV Action dataset is introduced, consisting of skeletal data for 10 actions.Evaluation on three datasets confirms better performance compared to other pose-based and depth-based methods. Human actions can be considered as a sequence of body poses over time, usually represented by coordinates corresponding to human skeleton models. Recently, a variety of low-cost devices have been released, able to produce markerless real time pose estimation. Nevertheless, limitations of the incorporated RGB-D sensors can produce inaccuracies, necessitating the utilization of alternative representation and classification schemes in order to boost performance. In this context, we propose a method for action recognition where skeletal data are initially processed in order to obtain robust and invariant pose representations and then vectors of dissimilarities to a set of prototype actions are computed. The task of recognition is performed in the dissimilarity space using sparse representation. A new publicly available dataset is introduced in this paper, created for evaluation purposes. The proposed method was also evaluated on other public datasets, and the results are compared to those of similar methods.

HEp-2 cells classification via sparse representation of textural features fused into dissimilarity space

Abstract Autoimmune diseases are proven to be connected with the occurrence of autoantibodies in patient serum. Antinuclear autoantibodies (ANAs) identification can be accomplished in a laboratory using indirect immunofluorescence (IIF) imaging. In this paper a system for automatic classification of staining patterns on HEp-2 fluorescence images is proposed. Our method utilizes two descriptors in order to encode gradient and textural characteristics of the depicted patterns. Along with distribution of SIFT features, we propose the new GoC-LBP descriptor based on co-occurrences of uniform Local Binary Patterns along directions dictated by the orientation of local gradient. At a second stage, the descriptors are fused while creating a dissimilarity representation of an image. A powerful classification scheme is incorporated, utilizing a discriminative sparse representation-based scheme for the classification. Experiments were conducted using a publicly available dataset, comparing the obtained performance to recently reported results of a relevant contest, demonstrating the effectives of the proposed method.

A framework for gait-based recognition using Kinect

Dynamic characteristics of gait are utilized for identity and gender recognition.A new publicly available dataset for gait recognition is presented.Our algorithm can operate extremely well with a small sample training size.The proposed framework follows a biologicaly inspired human motion analysis.The hierarchy of feature representations results in a high level description. Gait analysis has gained new impetus over the past few years. This is mostly due to the launch of low cost depth cameras accompanied with real time pose estimation algorithms. In this work we focus on the problem of human gait recognition. In particular, we propose a modification of a framework originally designed for the task of action recognition and apply it to gait recognition. The new scheme allows us to achieve complex representations of gait sequences and thus express efficiently the dynamic characteristics of human walking sequences. The representational power of the suggested model is evaluated on a publicly available dataset where we achieved up to 93.29% identification rate, 3.1% EER on the verification task and 99.11% gender recognition rate.

Gait-based gender recognition using pose information for real time applications

Biological cues inherent in human motion play an important role in the context of social communication. While recognizing the gender of other people is important for humans, security, advertisement and population statistics systems could also benefit from such kind of information. In this work for first time we propose a method suitable for real time gait based gender recognition relying on poses estimated from depth images. We provide evidence that pose based representation estimated by depth images could greatly benefit the problem of gait analysis. Given a gait sequence, in every frame the dynamics of gait motion are encoded using an angular representation. In particular several skeletal primitives are expressed as two Euler angles that cast votes into aggregated histograms. These histograms are then normalized, concatenated and projected onto a PCA basis in order to form the final sequence descriptor. We evaluated our method on a newly created dataset -UPCVgait - captured with Microsoft Kinect, consisting of 5 gait sequences performed by 30 subjects. An RBF kernel SVM used for classification in a leave one person out scheme on gait sequences of arbitrary length as well as on variable number of frames confirms the efficiency of our method.

Face recognition via local sparse coding

In this paper the face recognition problem is addressed in a part-based sparse approach through the comparison of respective facial regions between different images. To this purpose, a sparse coding procedure is applied to non-overlapping patches derived from frontal-face images, in order to extract local facial information. An adequate measure is introduced, incorporating the resulted sparse representation along with the Hamming distance, in order to express pairwise similarities between faces. Finally, a simple Nearest Neighbor classifier is employed to determine the identity of each facial image. In addition, a new criterion is presented for the rejection of outliers. The emerged face recognition scheme is evaluated using publicly available facial image databases, and the results are compared with those of other well-established recognition methods.

Gait based recognition via fusing information from Euclidean and Riemannian manifolds

Pose-based gait recognition using Euclidean and Riemannian feature representations.Euclidean representation is based on a residual aggregation method.Riemannian is based on the covariance representation of a sequence.A new publicly available dataset acquired using Kinect 2 is presented.Fusion and Classification is performed via SRC in RKHS. Gait is a particular periodical type of human motion with several unique characteristics for every person. In this work we focus on the problem of pose based gait recognition. The contribution of the proposed work is threefold. First we represent every gait sequence according to both the deviation of the poses from an appropriate global model, as well as the intra-sequence pose variability. Secondly, we propose a method which allows us to fuse information from feature representations from both Euclidean and Riemannian spaces by mapping data in a Reproducing Kernel Hilbert Space (RKHS). Classification is then performed using a kernelized version of the SRC algorithm. Third we present a new publicly available dataset for pose based gait recognition captured with Kinect V2. Experimental evaluation reveals state-of-the-art performance in both recognition and verification tasks and a capacity for real-time operation.

HEp-2 cells classification using locally aggregated features mapped in the dissimilarity space

Indirect Immunofluorescence (IIF) followed by manual evaluation of the acquired slides from specialized personnel is the preferred laboratory technique used for the detection of Antinucleolar Antibodies (ANAs) in patient serum. In this procedure, several limitations appear and thus several automatic techniques have been proposed for the task of ANA detection. In this paper we propose a system for automatic classification of HEp-2 staining patterns, inspired by a recently proposed method for aggregating local image (SIFT) features into a compact and fixed length representation. More specifically we present a novel framework in which aggregated features are mapped into feature vectors in the dissimilarity space where the dimensionality of the descriptors is “naturally reduced”. The final descriptor is low dimensional, while evaluation on a recently published dataset yields state of the art results.

Collaborative Sparse Representation in Dissimilarity Space for Classification of Visual Information

In this work we perform a thorough evaluation of the most popular CR-based classification scheme, the SRC, on the task of classification in dissimilarity space. We examine the performance utilizing a large set of public domain dissimilarity datasets mainly derived from classification problems relevant to visual information. We show that CR-based methods can exhibit remarkable performance in challenging situations characterized by extreme non-metric and non-Euclidean behavior, as well as limited number of available training samples per class. Furthermore, we investigate the structural qualities of a dataset necessitating the use of such classifiers. We demonstrate that CR-based methods have a clear advantage on dissimilarity data stemming from extended objects, manifold structures or a combination of these qualities. We also show that the induced sparsity during CR, is of great significance to the classification performance, especially in cases with small representative sets in the training data and large number of classes.

Parsimonious Coding and Verification of Offline Handwritten Signatures

A common practice for addressing the problem of verifying the presence, or the consent of a person in many transactions is to utilize the handwritten signature. Among others, the offline or static signature is a valuable tool in forensic related studies. Thus, the importance of verifying static handwritten signatures still poses a challenging task. Throughout the literature, gray-level images, composed of handwritten signature traces are subjected to numerous processing stages; their outcome is the mapping of any input signature image in a so-called corresponding feature space. Pattern recognition techniques utilize this feature space, usually as a binary verification problem. In this work, sparse dictionary learning and coding are for the first time employed as a means to provide a feature space for offline signature verification, which intuitively adapts to a small set of randomly selected genuine reference samples, thus making it attractable for forensic cases. In this context, the K-SVD dictionary learning algorithm is employed in order to create a writer oriented lexicon. For any signature sample, sparse representation with the use of the writers lexicon and the Orthogonal Matching Pursuit algorithm generates a weight matrix; features are then extracted by applying simple average pooling to the generated sparse codes. The performance of the proposed scheme is demonstrated using the popular CEDAR, MCYT75 and GPDS300 signature datasets, delivering state of the art results.

Pose-based gait recognition with local gradient descriptors and hierarchically aggregated residuals

Abstract. We focus on the problem of pose-based gait recognition. Our contribution is two-fold. First, we incorporate a local histogram descriptor that allows us to encode the trajectories of selected limbs via a one-dimensional version of histogram of oriented gradients features. In this way, a gait sequence is encoded into a sequence of local gradient descriptors. Second, we utilize a robust encoding method in which the residuals of local descriptors, with respect to a discriminative model, are aggregated into fixed length vectors. This technique combines the advantages of both residual aggregation and soft-assignment techniques, resulting in a powerful vector representation. For classification purposes, we use a nonlinear kernel to map vectors into a reproducing kernel Hilbert space. Then, we classify an encoded gait sequence according to the sparse representation-based classification method. Experimental evaluation on two publicly available datasets demonstrates the effectiveness of the proposed scheme on both recognition and verification tasks.