Si Chen

Online MIL tracking with instance-level semi-supervised learning

In this paper we propose an online multiple instance boosting algorithm with instance-level semi-supervised learning, termed SemiMILBoost, to achieve robust object tracking. Our work revisits the multiple instance learning (MIL) formulation to alleviate the drifting problem in tracking, which addresses two key issues in the existing MIL based tracking-by-detection methods, i.e., the unselective treatment of instances in the positive bag during weak classifier updating and the lack of object prior knowledge in instance modeling. We tackle both issues in a principled way by using a robust SemiMILBoost algorithm, which treats instances in the positive bag as unlabeled while the ones in the negative bag as negative. To improve the discriminability of weak classifiers online, we iteratively update them with the pseudo-labels and importance of all instances in the positive bag, which are predicted by employing the instance-level semi-supervised learning technique with object prior knowledge during boosting. Experimental results demonstrate that our proposed algorithm outperforms the state-of-the-art tracking methods on several challenging video sequences.

Discriminative local collaborative representation for online object tracking

Sparse representation has been widely applied to object tracking. However, most sparse representation based trackers only use the holistic template to encode the candidates, where the discriminative information to separate the target from the background is ignored. In addition, the sparsity assumption with the l1 norm minimization is computationally expensive. In this paper, we propose a robust discriminative local collaborative (DLC) representation algorithm for online object tracking. DLC collaboratively uses the local image patches of both the target templates and the background ones to encode the candidates by an efficient local regularized least square solver with the l2 norm minimization, where the feature vectors are obtained by employing an effective discriminative-pooling method. Furthermore, we formulate the tracking as a discriminative classification problem, where the classifier is online updated by using the candidates predicted according to the residuals of their local patches. To adapt to the appearance changes, we iteratively update the dictionary with the foreground and background templates from the current frame and take occlusions into account as well. Experimental results demonstrate that our proposed algorithm performs favorably against the state-of-the-art trackers on several challenging video sequences.

Robust visual tracking via online semi-supervised co-boosting

This paper proposes a novel visual tracking algorithm via online semi-supervised co-boosting, which investigates the benefits of co-boosting (i.e., the integration of co-training and boosting) and semi-supervised learning in the online tracking process. Existing discriminative tracking algorithms often use the classification results to update the classifier itself. However, the classification errors are easily accumulated during the self-training process. In this paper, we employ an effective online semi-supervised co-boosting framework to update the weak classifiers built on two different feature views. In this framework, the pseudo-label and importance of an unlabeled sample are estimated based on the additive logistic regression for an integration of a prior model and an online classifier learned on one feature view, and then used to update the weak classifiers built on the other feature view. The proposed algorithm has a good ability to recover from drifting by incorporating prior knowledge of the object while being adaptive to appearance changes by effectively combining the complementary strengths of different feature views. Experimental results on a series of challenging video sequences demonstrate the superior performance of our algorithm compared to state-of-the-art tracking algorithms.

Multi-label Learning Based Deep Transfer Neural Network for Facial Attribute Classification

Abstract Deep Neural Network (DNN) has recently achieved outstanding performance in a variety of computer vision tasks, including facial attribute classification. The great success of classifying facial attributes with DNN often relies on a massive amount of labelled data. However, in real-world applications, labelled data are only provided for some commonly used attributes (such as age, gender); whereas, unlabelled data are available for other attributes (such as attraction, hairline). To address the above problem, we propose a novel deep transfer neural network method based on multi-label learning for facial attribute classification, termed FMTNet, which consists of three sub-networks: the Face detection Network (FNet), the Multi-label learning Network (MNet) and the Transfer learning Network (TNet). Firstly, based on the Faster Region-based Convolutional Neural Network (Faster R-CNN), FNet is fine-tuned for face detection. Then, MNet is fine-tuned by FNet to predict multiple attributes with labelled data, where an effective loss weight scheme is developed to explicitly exploit the correlation between facial attributes based on attribute grouping. Finally, based on MNet, TNet is trained by taking advantage of unsupervised domain adaptation for unlabelled facial attribute classification. The three sub-networks are tightly coupled to perform effective facial attribute classification. A distinguishing characteristic of the proposed FMTNet method is that the three sub-networks (FNet, MNet and TNet) are constructed in a similar network structure. Extensive experimental results on challenging face datasets demonstrate the effectiveness of our proposed method compared with several state-of-the-art methods.

A novel robust model fitting approach towards multiple-structure data segmentation

A novel model fitting approach via Structure Decision Graph (SDG) is proposed.SDG is constructed based on the weight score and minimum arrived distance.P-distance is developed to measure similarity via continuous consensus set.SDG is less disturbed by noises and outliers and easy to implement.Experiments show the superiority of SDG for multiple-structure segmentation. We propose a novel and effective robust model fitting approach based on the Structure Decision Graph (SDG) to segment multiple-structure data in the presence of outliers. The proposed approach is motivated by the observations that each structure can be characterized by one representative hypothesis, called as the Structure Prototype (SP), and the SPs have relatively large distances among them. In this paper, instead of analyzing each hypothesis individually, the residuals over all the hypotheses are used to explicitly construct an SDG, where a sorted weight score set and a minimum arrived distance set are respectively computed. Based on the SDG, the SPs corresponding to different structures can be easily determined. Compared with conventional robust model fitting approaches, one distinguishing characteristic of our approach is that the clustering procedure is not required. Therefore, the proposed approach is less disturbed by noises and outliers, and is relatively easy to implement. Experimental results on synthetic data and real-world image datasets demonstrate the superiority of the proposed approach over the state-of-the-art robust model fitting approaches for multiple-structure data segmentation.

Expression-targeted feature learning for effective facial expression recognition

Abstract In this paper, we propose a novel expression-targeted feature learning (ETFL) method for effective facial expression recognition, which takes advantage of multi-task learning for discriminative feature learning. Specifically, the common features are firstly extracted from the lower layers of CNN. Then, based on the common features, the expression-specific features (ESF) are respectively learned for each facial expression via multi-task learning. In order to enhance the discriminability of ESF, we develop a joint loss (the combination of the center loss and a novel inter-class loss) to explicitly reduce intra-class variations while enlarging inter-class differences. Furthermore, we introduce the sample-sensitive weights and the soft-expression weights to balance the joint loss for better performance. Finally, all ESFs are combined for final classification. ETFL effectively exploits the relationship among all facial expressions, which leads to superiority feature discriminability. Experiments on public facial expression databases demonstrate the effectiveness of ETFL compared with several state-of-the-art methods.

Adaptive Metric Learning with the Low Rank Constraint

Good quality distance metrics can significantly promote the performance of many computer vision applications. In order to learn an appropriate distance metric, most of existing metric learning approaches restrict the learned distances between similar pairs to be smaller than a given lower bound, while the learned distances between dissimilar pairs are required to be larger than a given upper bound. However, the learned metrics may not perform well by leveraging the fixed bounds, especially when the data distributions are complex in practical applications. Besides, most methods attempt to learn a distance metric with a full rank matrix transformation from the given training data, which is not only inefficient to compute but also prone to overfitting. In this paper, we propose an Adaptive Metric Learning with the Low Rank Constraint (AML-LR) method, which restricts the learned distances between examples of pairs using adaptive bounds and meanwhile the rank of the learned matrix is minimized. Therefore, the learned metric is adaptive to different data distributions and robust to avoid overfitting. To solve the proposed optimization problem efficiently, we present an effective optimization algorithm based on the accelerated gradient method. Experimental results on UCI datasets and face verification databases demonstrate that AML-LR achieves competitive results compared with other state-of-the-art metric learning methods.