Search

This paper describes new feed-forward architectural and configural/holistic strategies for robust face recognition. This includes adaptive and robust correlation filters that lock on both appearance and location, and recognition-by-parts using boosting over strangeness driven weak learners. The utility of the proposed architectural strategies, shown with respect to different databases, includes occlusion, disguise, and temporal changes. The results obtained confirm and complement key findings on the ways people recognize each other, among them that the facial features are processed holistically and that the eyebrows are among the most important features for recognition. [ABSTRACT FROM AUTHOR]

In this paper, we propose a unified scheme of subspace and distance metric learning under the Bayesian framework for face recognition. According to the local distribution of data, we divide the k-nearest neighbors of each sample into the intra-person set and the inter-person set, and we aim to learn a distance metric in the embedding subspace, which can make the distances between the sample and its intra-person set smaller than the distances between it and its inter-person set. To reach this goal, we define two variables, that is, the intra-person distance and the inter-person distance, which are from two different probabilistic distributions, and we model the goal with minimizing the overlap between two distributions. Inspired by the Bayesian classification error estimation, we formulate it by minimizing the Bhattachyrra coefficient between two distributions. The power of the proposed approach are demonstrated by a series of experiments on the CMU-PIE face database and the extended YALE face database. [ABSTRACT FROM AUTHOR]

Pose variance remains a challenging problem for face recognition. In this paper, a scheme including image synthesis and recognition is proposed to improve the performance of automatic face recognition system. In the image synthesis part, a series of pose-variant images are produced based on three images respectively with front, left-profile, right-profile poses, and are added into the gallery in order to overcome the pose inconsistence between probes and images in the database. In the recognition part, a multi-level fusion method based on Gabor-combined features and gray-intensity features (GCGIF) is presented. Both amplitude features and phase features extracted through Gabor filters are utilized. Fusion is introduced in both the face representation level and the confidence level. Experiment results show that the integrated scheme achieve superior recognition performance. [ABSTRACT FROM AUTHOR]

Integral projections reduce the size of input data by transforming 2D images into significantly simpler 1D signals, while retaining useful information to solve important computer vision problems like object detection, location, and tracking. However, previous attempts typically rely on simple heuristic analysis such as searching for minima or maxima in the resulting projections. We introduce a more rigorous and formal modeling framework based on a small set of integral projections -thus, we will call them 1.5D models- and show that this model-based analysis overcomes many of the difficulties and limitations of alternative projection methods. The proposed approach proves to be particularly adequate for the specific domain of human face processing. The problems of face detection, facial feature location, and tracking in video sequences are studied under the unifying proposed framework. [ABSTRACT FROM AUTHOR]

A key problem for "face in the crowd" recognition from existing surveillance cameras in public spaces (such as mass transit centres) is the issue of pose mismatches between probe and gallery faces. In addition to accuracy, scalability is also important, necessarily limiting the complexity of face classification algorithms. In this paper we evaluate recent approaches to the recognition of faces at relatively large pose angles from a gallery of frontal images and propose novel adaptations as well as modifications. Specifically, we compare and contrast the accuracy, robustness and speed of an Active Appearance Model (AAM) based method (where realistic frontal faces are synthesized from non-frontal probe faces) against bag-of-features methods (which are local feature approaches based on block Discrete Cosine Transforms and Gaussian Mixture Models). We show a novel approach where the AAM based technique is sped up by directly obtaining pose-robust features, allowing the omission of the computationally expensive and artefact producing image synthesis step. Additionally, we adapt a histogram-based bag-of-features technique to face classification and contrast its properties to a previously proposed direct bag-of-features method. We also show that the two bag-of-features approaches can be considerably sped up, without a loss in classification accuracy, via an approximation of the exponential function. Experiments on the FERET and PIE databases suggest that the bag-of-features techniques generally attain better performance, with significantly lower computational loads. The histogram-based bag-of-features technique is capable of achieving an average recognition accuracy of 89% for pose angles of around 25 degrees. [ABSTRACT FROM AUTHOR]

Small sample size and severe facial variation are two challenging problems for face recognition. In this paper, we propose the SIS (Single Image Subspace) approach to address these two problems. To deal with the former one, we represent each single image as a subspace spanned by its synthesized (shifted) samples, and employ a newly designed subspace distance metric to measure the distance of subspaces. To deal with the latter one, we divide a face image into several regions, compute the contribution scores of the training samples based on the extracted subspaces in each region, and aggregate the scores of all the regions to yield the ultimate recognition result. Experiments on well-known face databases such as AR, Extended YALE and FERET show that the proposed approach outperforms some renowned methods not only in the scenario of one training sample per person, but also in the scenario of multiple training samples per person with significant facial variations. [ABSTRACT FROM AUTHOR]

Extending recognition to uncontrolled situations is a key challenge for practical face recognition systems. Finding efficient and discriminative facial appearance descriptors is crucial for this. Most existing approaches use features of just one type. Here we argue that robust recognition requires several different kinds of appearance information to be taken into account, suggesting the use of heterogeneous feature sets. We show that combining two of the most successful local face representations, Gabor wavelets and Local Binary Patterns (LBP), gives considerably better performance than either alone: they are complimentary in the sense that LBP captures small appearance details while Gabor features encode facial shape over a broader range of scales. Both feature sets are high dimensional so it is beneficial to use PCA to reduce the dimensionality prior to normalization and integration. The Kernel Discriminative Common Vector method is then applied to the combined feature vector to extract discriminant nonlinear features for recognition. The method is evaluated on several challenging face datasets including FRGC 1.0.4, FRGC 2.0.4 and FERET, with promising results. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel appearance-based representation, called Structured Ordinal Feature (SOF). SOF is a binary string encoded by combining eight ordinal blocks in a circle symmetrically. SOF is invariant to linear transformations on images and is flexible enough to represent different local structures of different complexity. We further extend SOF to Multi-scale Structured Ordinal Feature (MSOF) by concatenating binary strings of multi-scale SOFs at a fix position. In this way, MSOF encodes not only microstructure but also macrostructure of image patterns, thus provides a more powerful image representation. We also present an efficient algorithm for computing MSOF using integral images. Based on MSOF, statistical analysis and learning are performed to select most effective features and construct classifiers. The proposed method is evaluated with face recognition experiments, in which we achieve a high rank-1 recognition rate of 98.24% on FERET database. [ABSTRACT FROM AUTHOR]

Recognition in uncontrolled situations is one of the most important bottlenecks for practical face recognition systems. We address this by combining the strengths of robust illumination normalization, local texture based face representations and distance transform based matching metrics. Specifically, we make three main contributions: (i) we present a simple and efficient preprocessing chain that eliminates most of the effects of changing illumination while still preserving the essential appearance details that are needed for recognition; (ii) we introduce Local Ternary Patterns (LTP), a generalization of the Local Binary Pattern (LBP) local texture descriptor that is more discriminant and less sensitive to noise in uniform regions; and (iii) we show that replacing local histogramming with a local distance transform based similarity metric further improves the performance of LBP/LTP based face recognition. The resulting method gives state-of-the-art performance on three popular datasets chosen to test recognition under difficult illumination conditions: Face Recognition Grand Challenge version 1 experiment 4, Extended Yale-B, and CMU PIE. [ABSTRACT FROM AUTHOR]

In this paper we propose a novel boosting based classification algorithm, SODA-Boosting (where SODA stands for Second Order Discriminant Analysis). Unlike the conventional AdaBoost based algorithms widely applied in computer vision, SODA-Boosting does not involve time consuming procedures to search a huge feature pool in every iteration during the training stage. Instead, in each iteration SODA-Boosting efficiently computes discriminative weak classifiers in closed-form, based on reasonable hypotheses on the distribution of the weighted training samples. As an application, SODA-Boosting is employed for image based gender recognition. Experimental results on publicly available FERET database are reported. The proposed algorithm achieved accuracy comparable to state-of-the-art approaches, and demonstrated superior performance to relevant boosting based algorithms. [ABSTRACT FROM AUTHOR]

This paper describes a model-assisted system for reconstruction of 3D faces from a single consumer quality camera using a structure from motion approach. Typical multi-view stereo approaches use the motion of a sparse set of features to compute camera pose followed by a dense matching step to compute the final object structure. Accurate pose estimation depends upon precise identification and matching of feature points between images, but due to lack of texture on large areas of the face, matching is prone to errors. To deal with outliers in both the sparse and dense matching stages, previous work either relies on a strong prior model for face geometry or imposes restrictions on the camera motion. Strong prior models result in a serious compromise in final reconstruction quality and typically bear a signature resemblance to a generic or mean face. Model-based techniques, while giving the appearance of face detail, in fact carry this detail over from the model prior. Face features such as beards, moles, and other characteristic geometry are lost. Motion restrictions such as allowing only pure rotation are nearly impossible to satisfy by the end user, especially with a handheld camera. We significantly improve the robustness and flexibility of existing monocular face reconstruction techniques by introducing a deformable generic face model only at the pose estimation, face segmentation, and preprocessing stages. To preserve data fidelity in the final reconstruction, this generic model is discarded completely and dense matching outliers are removed using tensor voting: a purely data-driven technique. Results are shown from a complete end to end system. [ABSTRACT FROM AUTHOR]

This paper presents a patch-based approach for pose estimation from single images using a kernelized density voting scheme. We introduce a boosting-like algorithm that models the density using a mixture of weighted ‘weak' estimators. The ‘weak' density estimators and corresponding weights are learned iteratively from a training set, providing an efficient method for feature selection. Given a query image, voting is performed by reference patches similar in appearance to query image patches. Locality in the voting scheme allows us to handle occlusions and reduces the size of the training set required to cover the space of possible poses and appearance. Finally, the pose is estimated as the dominant mode in the density. Multimodality can be handled by looking at multiple dominant modes. Experiments carried out on face and articulated body pose databases show that our patch-based pose estimation algorithm generalizes well to unseen examples, is robust to occlusions and provides accurate pose estimation. [ABSTRACT FROM AUTHOR]

We present the first framework for detecting, localizing and classifying visual traits of object classes, e.g. gender or age of human faces, from arbitrary viewpoints. We embed all three tasks in a viewpoint-invariant model derived from local scale-invariant features (e.g. SIFT), where features are probabilistically quantified in terms of their occurrence, appearance, geometry and relationship to visual traits of interest. An appearance model is first learned for the object class, after which a Bayesian classifier is trained to identify the model features indicative of visual traits. The advantage of our framework is that it can be applied and evaluated in realistic scenarios, unlike other trait classification techniques that assume data that is single-viewpoint, pre-aligned and cropped from background distraction. Experimentation on the standard color FERET database shows our approach can automatically identify the visual cues in face images linked to the trait of gender. Combined detection, localization and gender classification error rates are a) 15% over a 180-degree range of face viewpoint and b) 13% in frontal faces, lower than other reported results. [ABSTRACT FROM AUTHOR]

3D face tracking is an important component for many computer vision applications. Most state-of-the-art tracking algorithms can be characterized as being either intensity- or feature-based. The intensity-based tracker relies on the brightness constraint while the feature-based tracker utilizes 2D local feature correspondences. In this paper, we propose a hybrid tracker for robust 3D face tracking. Instead of relying on single source of information, the hybrid tracker integrates feature correspondence and brightness constraints within a nonlinear optimization framework. The proposed method can track the 3D face pose reliably in real-time. We have conducted a series of evaluations to compare the performance of the proposed tracker with other state-of-the-art trackers. The experiments consist of synthetic sequences with simulation of different environmental factors, real sequences with estimated ground truth, and sequences from a real-world HCI application. The proposed tracker is shown to be superior in both accuracy and robustness. [ABSTRACT FROM AUTHOR]

Over the last two decades automatic facial expression recognition has become an active research area. Facial expressions are an important channel of non-verbal communication, and can provide cues to emotions and intentions. This paper introduces a novel method for facial expression recognition, by assembling contour fragments as discriminatory classifiers and boosting them to form a strong accurate classifier. Detection is fast as features are evaluated using an efficient lookup to a chamfer image, which weights the response of the feature. An Ensemble classification technique is presented using a voting scheme based on classifiers responses. The results of this research are a 6-class classifier (6 basic expressions of anger, joy, sadness, surprise, disgust and fear) which demonstrate competitive results achieving rates as high as 96% for some expressions. As classifiers are extremely fast to compute the approach operates at well above frame rate. We also demonstrate how a dedicated classifier can be consrtucted to give optimal automatic parameter selection of the detector, allowing real time operation on unconstrained video. [ABSTRACT FROM AUTHOR]

This paper introduces a novel method of single camera gait reconstruction which is independent of the walking direction and of the camera parameters. Recognizing people by gait has unique advantages with respect to other biometric techniques: the identification of the walking subject is completely unobtrusive and the identification can be achieved at distance. Recently much research has been conducted into the recognition of fronto-parallel gait. The proposed method relies on the very nature of walking to achieve the independence from walking direction. Three major assumptions have been done: human gait is cyclic; the distances between the bone joints are invariant during the execution of the movement; and the articulated leg motion is approximately planar, since almost all of the perceived motion is contained within a single limb swing plane. The method has been tested on several subjects walking freely along six different directions in a small enclosed area. The results show that recognition can be achieved without calibration and without dependence on view direction. The obtained results are particularly encouraging for future system development and for its application in real surveillance scenarios. [ABSTRACT FROM AUTHOR]

Recently skeletal motion data is obtained from the motion capture and is used for movie and sports. The movie production does not need the skeletal motion data but the body surface data. It is difficult to generate body surface data from only skeletal motion data because muscle deforms according to the skeletal motion. Muscle deformation occurs with arm and leg joint rotation. In this paper, we visualize body surface deformation based on the deformation mechanism that is applicable to human motion according to anatomy based modeling. We propose the method generating body surface by covering the skeletal muscles using a thin film based on the level set method. We demonstrate the effectiveness of the system through the generation of the movement of a body builder by using the proposed system. [ABSTRACT FROM AUTHOR]

This paper addresses an aspect of sign language (SL) recognition that has largely been overlooked in previous work and yet is integral to signed communication. It is the most comprehensive work to-date on recognizing complex variations in sign appearances due to grammatical processes (inflections) which systematically modulate the temporal and spatial dimensions of a root sign word to convey information in addition to lexical meaning. We propose a novel dynamic Bayesian network - the Multichannel Hierarchical Hidden Markov Model (MH-HMM)- as a modelling and recognition framework for continuously signed sentences that include modulated signs. This models the hierarchical, sequential and parallel organization in signing while requiring synchronization between parallel data streams at sign boundaries. Experimental results using particle filtering for decoding demonstrate the feasibility of using the MH-HMM for recognizing inflected signs in continuous sentences. [ABSTRACT FROM AUTHOR]

We propose a novel method for 3D head reconstruction and view-invariant recognition from single 2D images. We employ a deterministic Shape From Shading (SFS) method with initial conditions estimated by Hybrid Principal Component Analysis (HPCA) and multi-level global optimization with error-dependent smoothness and integrability constraints. Our HPCA algorithm provides initial estimates of 3D range mapping for the SFS optimization, which is quite accurate and yields much improved 3D head reconstruction. The paper also includes significant contributions in novel approaches to global optimization and in SFS handling of variable and unknown surface albedo, a problem with unsatisfactory solutions by prevalent SFS methods. In the experiments, we reconstruct 3D head range images from 2D single images in different views. The 3D reconstructions are then used to recognize stored model persons. Empirical results show that our HPCA based SFS method provides 3D head reconstructions that notably improve the accuracy compared to other approaches. 3D reconstructions derived from side view (profile) images of 40 persons are tested against 80 3D head models and a recognition rate of over 90% is achieved. Such a capability was not demonstrated by any other method we are aware of. [ABSTRACT FROM AUTHOR]

This paper addresses the recovery of face models from stereo pairs of images in the presence of foreign-body occlusions. In the proposed approach, a 3D morphable model (3DMM) for faces is augmented by an occlusion map defined on the model shape, and occlusion is detected with minimal computational overhead by incorporating robust estimators in the fitting process. Additionally, the method uses an explicit model for texture (or reflectance) in addition to shape, which is in contrast to most existing multi-view methods that use a shape model alone. We argue that both model components are required to handle certain classes of occluders, and we present empirical results to support this claim. In fact, the empirical results in this paper suggest that even in the absence of occlusions, stereo reconstruction using existing shape-only face models can perform poorly by some measures, and that the inclusion of an explicit texture model may be worth its computational expense. [ABSTRACT FROM AUTHOR]

In tracking face and facial actions of unknown people, it is essential to take into account two components of facial shape variations: shape variation between people and variation caused by different facial actions such as facial expressions. This paper presents a monocular method of tracking faces and facial actions using a multilinear face model that treats interpersonal and intrapersonal shape variations separately. We created this method using a multilinear face model by integrating two different frameworks: particle filter-based tracking for time-dependent facial action and pose estimation and incremental bundle adjustment for person-dependent shape estimation. This unique combination together with multilinear face models is the key to tracking faces and facial actions of arbitrary people in real time with no pre-learned individual face models. Experiments using real video sequences demonstrate the effectiveness of our method. [ABSTRACT FROM AUTHOR]

In this paper, we present an effective approach for spatiotemporal face recognition from videos using an Extended set of Volume LBP (Local Binary Pattern features) and a boosting scheme. Among the key properties of our approach are: (1) the use of local Extended Volume LBP based spatiotemporal description instead of the holistic representations commonly used in previous works; (2) the selection of only personal specific facial dynamics while discarding the intra-personal temporal information; and (3) the incorporation of the contribution of each local spatiotemporal information. To the best of our knowledge, this is the first work addressing the issue of learning the personal specific facial dynamics for face recognition. We experimented with three different publicly available video face databases (MoBo, CRIM and Honda/UCSD) and considered five benchmark methods (PCA, LDA, LBP, HMMs and ARMA) for comparison. Our extensive experimental analysis clearly assessed the excellent performance of the proposed approach, significantly outperforming the comparative methods and thus advancing the state-of-the-art. [ABSTRACT FROM AUTHOR]

This paper presents a new method for face modeling and face recognition from a pair of calibrated stereo cameras. In a first step, the algorithm builds a stereo reconstruction of the face by adjusting the global transformation parameters and the shape parameters of a 3D morphable face model. The adjustment of the parameters is such that stereo correspondence between both images is established, i.e. such that the 3D-vertices of the model project on similarly colored pixels in both images. In a second step, the texture information is extracted from the image pair and represented in the texture space of the morphable face model. The resulting shape and texture coefficients form a person specific feature vector and face recognition is performed by comparing query vectors with stored vectors. To validate our algorithm, an extensive image database was built. It consists of stereo-pairs of 70 subjects. For recognition testing, the subjects were recorded under 6 different head directions, ranging from a frontal to a profile view. The face recognition results are very good, with 100% recognition on frontal views and 97% recognition on half-profile views. [ABSTRACT FROM AUTHOR]