Supervised Descent Method (SDM) has shown good performance in solving non-linear least squares problems in computer vision, giving state of the art results for the problem of face alignment. However, when SDM learns the generic descent maps, it is very difficult to avoid over-fitting due to the high dimensionality of the input features. In this paper we propose a Random Subspace SDM (RSSDM) that maintains the high accuracy on the training data and improves the generalization accuracy. Instead of using all the features for descent learning at each iteration, we randomly select sub-sets of the features and learn an ensemble of descent maps in the corresponding subspaces, one in each subspace. Then, we average the ensemble of descents to calculate the update of the iteration. We test the proposed methods on two representative regression problems, namely, 3D pose estimation and face alignment and show that RSSDM consistently outperforms SDM in both tasks in terms of accuracy (e.g., RSSDM is able to localize 4% more landmarks at error level of 0.1 on the challenging iBug dataset). RSSDM also holds several useful generalization properties: 1) it is more effective when the number of training samples is small-with 3 Monte-Carlo permutations RSSDM can achieve similar performance to SDM with 9 Monte-Carlo permutations; 2) it is less sensitive to the changes of the strength of the regularization-when the regularization parameter is changed to 10 times larger, the mean error increases 9.0% for SDM vs. 3.4% for RSSDM.

We propose an adaptive Bayesian hidden Markov model for fully unsupervised part-of-speech (POS) induction. The proposed model with its inference algorithm has two extensions to the first-order Bayesian HMM with Dirichlet priors. First our algorithm infers the optimal number of hidden states from the training corpus rather than fixes the dimensionality of state space beforehand. The second extension studies the Chinese unknown word processing module which measures similarities from both morphological properties and context distribution. Experimental results showed that both of these two extensions can help to find the optimal categories for Chinese in terms of both unsupervised clustering metrics and grammar induction accuracies on the Chinese Treebank.

Distribution calibration plays an important role in cross-domain learning. However, existing distribution distance metrics are not geodesic; therefore, they cannot measure the intrinsic distance between two distributions. In this paper, we calibrate two distributions by using the geodesic distance in Riemannian symmetric space. Our method learns a latent subspace in the reproducing kernel Hilbert space, where the geodesic distance between the distribution of the source and the target domains is minimized. The corresponding geodesic distance is thus equivalent to the geodesic distance between two symmetric positive definite (SPD) matrices defined in the Riemannian symmetric space. These two SPD matrices parameterize the marginal distributions of the source and target domains in the latent subspace. We carefully design an evolutionary algorithm to find a local optimal solution that minimizes this geodesic distance. Empirical studies on face recognition, text categorization, and web image annotation suggest the effectiveness of the proposed scheme.

Dimension reduction algorithms have attracted a lot of attentions in face recognition because they can select a subset of effective and efficient discriminative features in the face images. Most of dimension reduction algorithms can not well model both the intra-class geometry and interclass discrimination simultaneously. In this paper, we introduce the Discriminative Hessian Eigenmaps (DHE), a novel dimension reduction algorithm to address this problem. DHE will consider encoding the geometric and discriminative information in a local patch by improved Hessian Eigenmaps and margin maximization respectively. Empirical studies on public face database thoroughly demonstrate that DHE is superior to popular algorithms for dimension reduction, e.g., FLDA, LPP, MFA and DLA.

Near-duplicate video retrieval is becoming more and more important with the exponential growth of the Web. Though various approaches have been proposed to address this problem, they are mainly focusing on the retrieval accuracy while infeasible to query on Web scale video database in real time. This paper proposes a novel method to address the efficiency and scalability issues for near-duplicate We video retrieval. We introduce a compact spatiotemporal feature to represent videos and construct an efficient data structure to index the feature to achieve real-time retrieving performance. This novel feature leverages relative gray-level intensity distribution within a frame and temporal structure of videos along frame sequence. The new index structure is proposed based on inverted file to allow for fast histogram intersection computation between videos. To demonstrate the effectiveness and efficiency of the proposed methods we evaluate its performance on an open Web video data set containing about 10K videos and compare it with four existing methods in terms of precision and time complexity. We also test our method on a data set containing about 50K videos and 11M key-frames. It takes on average 17ms to execute a query against the whole 50K Web video data set.

Is it possible to train a learning model to separate tigers from elks when we have 1) labeled samples of leopard and zebra and 2) unlabelled samples of tiger and elk at hand? Cross-domain learning algorithms can be used to solve the above problem. However, existing cross-domain algorithms cannot be applied for dimension reduction, which plays a key role in computer vision tasks, e.g., face recognition and web image annotation. This paper envisions the cross-domain discriminative dimension reduction to provide an effective solution for cross-domain dimension reduction. In particular, we propose the cross-domain discriminative Hessian Eigenmaps or CDHE for short. CDHE connects training and test samples by minimizing the quadratic distance between the distribution of the training set and that of the test set. Therefore, a common subspace for data representation can be well preserved. Furthermore, we basically expect the discriminative information used to separate leopards and zebra can be shared to separate tigers and elks, and thus we have a chance to duly address the above question. Margin maximization principle is adopted in CDHE so the discriminative information for separating different classes (e.g., leopard and zebra here) can be well preserved. Finally, CDHE encodes the local geometry of each training class (e.g., leopard and zebra here) in the local tangent space which is locally isometric to the data manifold and thus CDHE preserves the intraclass local geometry. The objective function of CDHE is not convex, so the gradient descent strategy can only find a local optimal solution. In this paper, we carefully design an evolutionary search strategy to find a better solution of CDHE. Experimental evidence on both synthetic and real word image datasets demonstrates the effectiveness of CDHE for cross-domain web image annotation and face recognition.

This paper presents a nonparametric discriminant HMM and applies it to facial expression recognition. In the proposed HMM, we introduce an effective nonparametric output probability estimation method to increase the discrimination ability at both hidden state level and class level. The proposed method uses a nonparametric adaptive kernel to utilize information from all classes and improve the discrimination at class level. The discrimination between hidden states is increased by defining membership coefficients which associate each reference vector with hidden states. The adaption of such coefficients is obtained by the Expectation Maximization (EM) method. Furthermore, we present a general formula for the estimation of output probability, which provides a way to develop new HMMs. Finally, we evaluate the performance of the proposed method on the CMU expression database and compare it with other nonparametric HMMs.

We present a Temporal Exemplar-based Bayesian Networks (TEBNs) for facial expression recognition. The proposed Bayesian Networks (BNs) consists of three layers: Observation layer, Exemplars layer and Prior Knowledge layer. In the Exemplars layer, exemplar-based model is integrated with BNs to improve the accuracy of probability estimation. In the Prior Knowledge layer, static BNs is extended to Temporal BNs by considering historical observations to model temporal behavior of facial expression. Experiment on CMU expression database illustrates that the proposed TEBNs is very efficient in modeling the evolution of facial deformation.

In this article, we propose a new postprocessing strategy, word suggestion, based on a multiple word trigger-pair language model for Chinese character recognizers. With the word suggestion strategy, Chinese character recognizers may even achieve a recognition rate greater than the top-n candidate recognition rate. To construct the multiple word trigger-pair model, data mining techniques are used to alleviate the intensive computation problem. Furthermore, rough set theory is first used in the study to discover negatively correlated relationships between words in order to prevent introducing wrong words in the process of word suggestion.

A good language model is essential to a postprocessing algorithm for recognition systems. Trigger pair model has been used to investigate long distance dependent relationship. However, previous trigger pair model has only one word for its trigger. It is desirable that more words can be observed in the trigger for a better prediction of the triggered word. In this work, we view establishing trigger pair model as mining association rules in a large database and create a multiple words trigger pair model by using an adapted A priori algorithm. The new trigger pair model can be used in the stage of finding best path from a word lattice as traditional trigger pair model can. Specially, it can be used to correct mistakes remaining in the final result as well. Those mistakes would be unavoidable for other language models.

Fuzzy-Attribute Graph (FAG) was proposed to handle fuzziness in the pattern primitives in structural pattern recognition. FAG has the advantage that we can combine several possible definitions into a single template, and hence only one matching is required instead of one for each definition. Also, each vertex or edge of the graph can contain fuzzy attributes to model real-life situations. However, in our previous approach, we need a human expert to define the templates for the fuzzy graph matching. This is usually tedious, time-consuming and error-prone. In this paper, we propose a learning algorithm that will, from a number of fuzzy examples, each of them being a FAG, find the smallest template that can be matched to the given patterns with respect to the matching metric.

High-dimensional grammars such as web grammars and plex grammars were used in syntactic recognition of complex 2-D or 3-D objects. In this paper, we present a simple modification, borrowing the concept of guards from concurrent programming to attributed grammar proposed by D. E. Knuth. We show that the resultant grammar can handle patterns described by the high-dimensional grammars. The only problem is that we may not have a simple ordering of the terminal symbols or pattern primitives. In some applications, such as on-line character recognition, the problem does not exist and hence presents a good candidate for the application. We also discuss the incorporation of fuzzy attributes and the necessary modification is hence introduced. Finally, the error transformations proposed by K. S. Fu can easily be taken into consideration and a powerful yet simple scheme presented.