Key Laboratory of Machine Perception
Key Laboratory of Machine Perception
Li J.,Beijing Jiaotong University |
Bai C.,Capital Normal University |
Lin Z.,Key Laboratory of Machine Perception |
Lin Z.,Shanghai JiaoTong University |
Yu J.,Beijing Jiaotong University
IEEE Transactions on Image Processing | Year: 2017
Demosaicking is the problem of reconstructing a color image from the raw image captured by a digital color camera that covers its only imaging sensor with a color filter array (CFA). Sparse representation-based demosaicking has been shown to produce superior reconstruction quality. However, almost all existing algorithms in this category use the CFAs, which are not specifically optimized for the algorithms. In this paper, we consider optimally designing CFAs for sparse representation-based demosaicking, where the dictionary is well-chosen. The fact that CFAs correspond to the projection matrices used in compressed sensing inspires us to optimize CFAs via minimizing the mutual coherence. This is more challenging than that for traditional projection matrices because CFAs have physical realizability constraints. However, most of the existing methods for minimizing the mutual coherence require that the projection matrices should be unconstrained, making them inapplicable for designing CFAs. We consider directly minimizing the mutual coherence with the CFA's physical realizability constraints as a generalized fractional programming problem, which needs to find sufficiently accurate solutions to a sequence of nonconvex nonsmooth minimization problems. We adapt the redistributed proximal bundle method to address this issue. Experiments on benchmark images testify to the superiority of the proposed method. In particular, we show that a simple sparse representation-based demosaicking algorithm with our specifically optimized CFA can outperform LSSC . To the best of our knowledge, it is the first sparse representation-based demosaicking algorithm that beats LSSC in terms of CPSNR. © 2016 IEEE.
He D.,Beijing Key Laboratory of Behavior and Mental Health |
He D.,Key Laboratory of Machine Perception |
He D.,Peking Tsinghua Center for Life science |
He D.,Peking University |
And 6 more authors.
Journal of Vision | Year: 2017
Saccadic eye movements cause rapid and dramatic displacements of the retinal image of the visual world, yet our conscious perception of the world remains stable and continuous. A popular explanation for this remarkable ability of our visual system to compensate for the displacements is the predictive feature remapping theory. The theory proposes that, before saccades, the representation of a visual stimulus can be predictively transferred from neurons that initially encode the stimulus to neurons whose receptive fields will encompass the stimulus location after the saccade. Visual adaptation aftereffect experiments performed by Melcher (2007) provided psychophysical evidence for this theory. However, it was argued that the visual aftereffects were not measured at the ''appropriate'' remapped location (Rolfs, Jonikaitis, Deubel, & Cavanagh, 2011). Therefore, whether the remapped representation contains feature information (e.g., orientation, motion direction, or contrast) is still a subject of intense debate. Here, to explore the nature of the predictive transfer during trans-saccadic perception, we measured visual aftereffects (tilt aftereffect, motion aftereffect, and threshold elevation aftereffect) at the appropriate remapped location of adapting stimuli before saccades. We observed a significant tilt aftereffect and motion aftereffect, but little threshold elevation aftereffect. Furthermore, the tilt aftereffect and motion aftereffect exhibited spatial specificity. These findings provide strong evidence for the predictive feature remapping theory and suggest that intermediate visual processing stages (i.e., extrastriate visual cortex) might be critical for feature remapping. Finally, we propose that the feature remapping process might also contribute to the spatiotopic representation of visual features. © Copyright 2017 The Authors.
Hu J.,Peking University |
Blue P.R.,Peking University |
Yu H.,Peking University |
Gong X.,Tongji University |
And 5 more authors.
Social Cognitive and Affective Neuroscience | Year: 2015
In human society, which is organized by social hierarchies, resources are usually allocated unequally and based on social status. In this study, we analyze how being endowed with different social statuses in a math competition affects the perception of fairness during asset allocation in a subsequent Ultimatum Game (UG). Behavioral data showed that when participants were in high status, they were more likely to reject unfair UG offers than in low status. This effect of social status correlated with activity in the right anterior insula (rAI) and with the functional connectivity between the rAI and a region in the anterior middle cingulate cortex, indicating that these two brain regions are crucial for integrating contextual factors and social norms during fairness perception. Additionally, there was an interaction between social status and UG offer fairness in the amygdala and thalamus, implicating the role of these regions in the modulation of social status on fairness perception. These results demonstrate the effect of social status on fairness perception and the potential neural underpinnings for this effect. © The Author (2015). Published by Oxford University Press.
Hong S.,Key Laboratory of Machine Perception |
Hong S.,Peking University |
Wu M.,Key Laboratory of Machine Perception |
Wu M.,Peking University |
And 3 more authors.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) | Year: 2017
Sequential data containing series of events with timestamps is commonly used to record status of things in all aspects of life, and is referred to as temporal event sequences. Learning vector representations is a fundamental task of temporal event sequence mining as it is inevitable for further analysis. Temporal event sequences differ from symbol sequences and numerical time series in that each entry is along with a corresponding time stamp and that the entries are usually sparse in time. Therefore, methods either on symbolic sequences such as word2vec, or on numerical time series such as pattern discovery perform unsatisfactorily. In this paper, we propose an algorithm called event2vec that solves these problems. We first present Event Connection Graph to summarize events while taking time into consideration. Then, we conducts a training Sample Generator to get clean and endless data. Finally, we feed these data to embedding neural network to get learned vectors. Experiments on real temporal event sequence data in medical area demonstrate the effectiveness and efficiency of the proposed method. The procedure is totally unsupervised without the help of expert knowledge. Thus can be used to improve the quality of health-care without any additional burden. © Springer International Publishing AG 2017.
Luo D.,Key Laboratory of Machine Perception |
Luo D.,Peking University |
Ding Y.,Key Laboratory of Machine Perception |
Ding Y.,Peking University |
And 5 more authors.
2014 IEEE International Conference on Mechatronics and Automation, IEEE ICMA 2014 | Year: 2014
Stand-up motion is among the most essential behaviors for humanoid robots. For achieving stable stand-up behavior, the traditional key-frame based motion planning methods are time-exhausted and expert knowledge dependent. On the other hand, classic trial-and-error based learning methods are inefficient due to the high degrees of freedom (DOFs) for humanoid robots and the difficulty in fixing appropriate reward functions. In this paper, a multi-stage learning approach is proposed to address the above issues. At the first stage, under a trajectory based motion control model, key motion frames sampled from human motion capture data (HMCD) are used for model initialization, through which the solution space could be pruned. At the second stage, the design of experiments (DOE) technique is introduced for fast and active searching in the pruned solution space. At the last stage, a refining process that adopts a stochastic gradient learning strategy is performed to achieve the final behavior. Under this three-stage learning framework, along with a simple heuristic reward function, the learning of the stand-up behavior for a kid-size humanoid robot is fulfilled successfully and efficiently. © 2014 IEEE.
Wang Z.,Peking University |
Yuan X.,Key Laboratory of Machine Perception |
Yuan X.,Peking University
2014 International Conference on Big Data and Smart Computing, BIGCOMP 2014 | Year: 2014
In this paper, we propose using timelines for 2D trajectory comparison. Trajectories directly rendered on a map do not show temporal information well, and are cluttered and unaligned. This make them difficult to compare. We convert trajectories to timelines, which naturally shows time, and are more compact and easy to align. In addition to simply showing how an attribute varies along the time, we further propose some novel timelines to show spatial-temporal features. We provide some use cases to show the benefit of our method. © 2014 IEEE.
Yin N.,Peking University |
Yin N.,Key Laboratory of Machine Perception |
Wang S.,Peking University |
Wang S.,Key Laboratory of Machine Perception |
And 4 more authors.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) | Year: 2014
In many applications, the data in time series appears highly periodic, but never exactly repeats itself. Such series are called pseudo periodic time series. The prediction of pseudo periodic time series is an important and nontrivial problem. Since the period interval is not fixed and unpredictable, errors will accumulate when traditional periodic methods are employed. Meanwhile, many time series contain a vast number of abnormal variations. These variations can neither be simply filtered out nor predicted by its neighboring points. Given that no specific method is available for pseudo periodic time series as of yet, the paper proposes a segment-wise method for the prediction of pseudo periodic time series with abnormal variations. Time series are segmented by the variation patterns of each period in the method. Only the segment corresponding to the target time series is chosen for prediction, which leads to the reduction of input variables. At the same time, the choice of the value highly correlated to the points-to-be-predicted enhances the prediction precision. Experimental results produced using data sets of China Mobile and biomedical signals both prove the effectiveness of the segment-wise method in improving the prediction accuracy of the pseudo periodic time series. © Springer International Publishing Switzerland 2014.
Zhang H.-Y.,Key Laboratory of Machine Perception |
Zhang H.-Y.,Peking University |
Wang L.-W.,Key Laboratory of Machine Perception |
Wang L.-W.,Peking University |
And 2 more authors.
Ruan Jian Xue Bao/Journal of Software | Year: 2013
Probabilistic graphical models are powerful tools for compactly representing complex probability distributions, efficiently computing (approximate) marginal and conditional distributions, and conveniently learning parameters and hyperparameters in probabilistic models. As a result, they have been widely used in applications that require some sort of automated probabilistic reasoning, such as computer vision and natural language processing, as a formal approach to deal with uncertainty. This paper surveys the basic concepts and key results of representation, inference and learning in probabilistic graphical models, and demonstrates their uses in two important probabilistic models. It also reviews some recent advances in speeding up classic approximate inference algorithms, followed by a discussion of promising research directions. ©Copyright 2013, Institute of Software, the Chinese Academy of Sciences.
Liu Z.-Q.,Peking University |
Liu Z.-Q.,Key Laboratory of Machine Perception |
Li H.-Y.,Peking University |
Li H.-Y.,Key Laboratory of Machine Perception |
And 4 more authors.
Jisuanji Xuebao/Chinese Journal of Computers | Year: 2010
Nowadays, the process-driven method for information system construction is more and more widely used. In the process-driven method, the process model has significantly influence on the data model. Unfortunately, the existing data model anomalies detection methods are only about the anomalies of data model itself. These methods don't take the process model into account. Similarly, the process model verification methods also lack for consideration of data model. This paper analyses those anomalies and gives a basic classification of three classes of them. This paper also proposes a model called Data-process Graph (DP-graph) to build up linkage between the data model and process model. Based on DP-Graph, DPGT method is proposed to detect anomalies of data model for business process model. Experimental results show the high detection rate of the anomalies of the method.
Luo Y.,Key Laboratory of Machine Perception |
Liu T.,Intelligent Systems Technology, Inc. |
Tao D.,Intelligent Systems Technology, Inc. |
Xu C.,Key Laboratory of Machine Perception
IEEE Transactions on Image Processing | Year: 2014
Distance metric learning (DML) is a critical factor for image analysis and pattern recognition. To learn a robust distance metric for a target task, we need abundant side information (i.e., the similarity/dissimilarity pairwise constraints over the labeled data), which is usually unavailable in practice due to the high labeling cost. This paper considers the transfer learning setting by exploiting the large quantity of side information from certain related, but different source tasks to help with target metric learning (with only a little side information). The state-of-the-art metric learning algorithms usually fail in this setting because the data distributions of the source task and target task are often quite different. We address this problem by assuming that the target distance metric lies in the space spanned by the eigenvectors of the source metrics (or other randomly generated bases). The target metric is represented as a combination of the base metrics, which are computed using the decomposed components of the source metrics (or simply a set of random bases); we call the proposed method, decomposition-based transfer DML (DTDML). In particular, DTDML learns a sparse combination of the base metrics to construct the target metric by forcing the target metric to be close to an integration of the source metrics. The main advantage of the proposed method compared with existing transfer metric learning approaches is that we directly learn the base metric coefficients instead of the target metric. To this end, far fewer variables need to be learned. We therefore obtain more reliable solutions given the limited side information and the optimization tends to be faster. Experiments on the popular handwritten image (digit, letter) classification and challenge natural image annotation tasks demonstrate the effectiveness of the proposed method. © 1992-2012 IEEE.