OpenCL is an open heterogeneous programming framework. Although OpenCL programs are functionally portable, they do not provide performance portability, so code transformation often plays an irreplaceable role. When adapting GPU-specific OpenCL kernels to run on multi-core/many-core CPUs, coarsening the thread granularity is necessary and thus has been extensively used. However, locality concerns exposed in GPU-specific OpenCL code are usually inherited without analysis, which may give side-effects on the CPU performance. Typically, the use of OpenCL’s local memory on multi-core/many-core CPUs may lead to an opposite performance effect, because local-memory arrays no longer match well with the hardware and the associated synchronizations are costly. To solve this dilemma, we actively analyze the memory access patterns using array-access descriptors derived from GPU-specific kernels, which can thus be adapted for CPUs by (1) removing all the unwanted local-memory arrays together with the obsolete barrier statements and (2) optimizing the coalesced kernel code with vectorization and locality re-exploitation. Moreover, we have developed an automated tool chain that makes this transformation of GPU-specific OpenCL kernels into a CPU-friendly form, which is accompanied with a scheduler that forms a new OpenCL runtime. Experiments show that the automated transformation can improve OpenCL kernel performance on a multi-core CPU by an average factor of 3.24. Satisfactory performance improvements are also achieved on Intel’s many-integrated-core coprocessor. The resultant performance on both architectures is better than or comparable with the corresponding OpenMP performance.
Human action recognition is currently one of the most active research areas in computer vision. It has been widely used in many applications, such as intelligent surveillance, perceptual interface, and content-based video retrieval. However, some extrinsic factors are barriers for the development of action recognition; e.g., human actions may be observed from arbitrary camera viewpoints in realistic scene. Thus, view-invariant analysis becomes important for action recognition algorithms, and a number of researchers have paid much attention to this issue. In this paper, we present a multi-view learning approach to recognize human actions from different views. As most existing multi-view learning algorithms often suffer from the problem of lacking data adaptiveness in the nearest neighborhood graph construction procedure, a robust locally adaptive multi-view learning algorithm based on learning multiple local L1-graphs is proposed. Moreover, an efficient iterative optimization method is proposed to solve the proposed objective function. Experiments on three public view-invariant action recognition datasets, i.e., ViHASi, IXMAS, and WVU, demonstrate data adaptiveness, effectiveness, and efficiency of our algorithm. More importantly, when the feature dimension is correctly selected (i.e.,>60), the proposed algorithm stably outperforms state-of-the-art counterparts and obtains about 6% improvement in recognition accuracy on the three datasets.
Head pose estimation has been considered an important and challenging task in computer vision. In this paper we propose a novel method to estimate head pose based on a deep convolutional neural network (DCNN) for 2D face images. We design an effective and simple method to roughly crop the face from the input image, maintaining the individual-relative facial features ratio. The method can be used in various poses. Then two convolutional neural networks are set up to train the head pose classifier and then compared with each other. The simpler one has six layers. It performs well on seven yaw poses but is somewhat unsatisfactory when mixed in two pitch poses. The other has eight layers and more pixels in input layers. It has better performance on more poses and more training samples. Before training the network, two reasonable strategies including shift and zoom are executed to prepare training samples. Finally, feature extraction filters are optimized together with the weight of the classification component through training, to minimize the classification error. Our method has been evaluated on the CAS-PEAL-R1, CMU PIE, and CUBIC FacePix databases. It has better performance than state-of-the-art methods for head pose estimation.
Named entity recognition (NER) is a core component in many natural language processing applications. Most NER systems rely on supervised machine learning methods, which depend on time-consuming and expensive annotations in different languages and domains. This paper presents a method for automatically building silver-standard NER corpora from Chinese Wikipedia. We refine novel and language-dependent features by exploiting the text and structure of Chinese Wikipedia. To reduce tagging errors caused by entity classification, we design four types of heuristic rules based on the characteristics of Chinese Wikipedia and train a supervised NE classifier, and a combined method is used to improve the precision and coverage. Then, we realize type identification of implicit mention by using boundary information of outgoing links. By selecting the sentences related with the domains of test data, we can train better NER models. In the experiments, large-scale NER corpora containing 2.3 million sentences are built from Chinese Wikipedia. The results show the effectiveness of automatically annotated corpora, and the trained NER models achieve the best performance when combining our silver-standard corpora with gold-standard corpora.
A hybrid power transmission system (HPTS) is a promising way to save energy in a hydraulic excavator and the electric machine is one of the key components of the system. In this paper, a design process for permanent-magnet synchronous machines (PMSMs) in a hybrid hydraulic excavator (HHE) is presented based on the analysis of the working conditions and requirements of an HHE. A parameterized design approach, which combines the analytical model and the 2D finite element method (FEM), is applied to the electric machine to improve the design efficiency and accuracy. The analytical model is employed to optimize the electric machine efficiency and obtain the stator dimension and flux density distribution. The rotor is designed with the FEM to satisfy the flux requirements obtained in stator design. The rotor configuration of the PMSM employs an interior magnet structure, thus resulting in some inverse saliency, which allows for much higher values in magnetic flux density. To reduce the rotor leakage, a disconnected type silicon steel block structure is adopted. To improve the air gap flux density distribution, the trapezoid permanent magnet (PM) and centrifugal rotor structure are applied to PMSM. Demagnetization and armature reactions are also taken into consideration and calculated by the FEM. A prototype of the newly designed electric machine has been fabricated and tested on the experimental platform. The analytical design results are validated by measurements.
Resampling is a critical procedure that is of both theoretical and practical significance for efficient implementation of the particle filter. To gain an insight of the resampling process and the filter, this paper contributes in three further respects as a sequel to the tutorial (Li et al., 2015). First, identical distribution (ID) is established as a general principle for the resampling design, which requires the distribution of particles before and after resampling to be statistically identical. Three consistent metrics including the (symmetrical) Kullback-Leibler divergence, Kolmogorov-Smirnov statistic, and the sampling variance are introduced for assessment of the ID attribute of resampling, and a corresponding, qualitative ID analysis of representative resampling methods is given. Second, a novel resampling scheme that obtains the optimal ID attribute in the sense of minimum sampling variance is proposed. Third, more than a dozen typical resampling methods are compared via simulations in terms of sample size variation, sampling variance, computing speed, and estimation accuracy. These form a more comprehensive understanding of the algorithm, providing solid guidelines for either selection of existing resampling methods or new implementations.
To address the problem of maneuvering target tracking, where the target trajectory has prolonged smooth regions and abrupt maneuvering regions, a modified variable rate particle filter (MVRPF) is proposed. First, a Cartesian-coordinate based variable rate model is presented. Compared with conventional variable rate models, the proposed model does not need any prior knowledge of target mass or external forces. Consequently, it is more convenient in practical tracking applications. Second, a maneuvering detection strategy is adopted to adaptively adjust the parameters in MVRPF, which helps allocate more state points at high maneuver regions and fewer at smooth regions. Third, in the presence of small measurement errors, the unscented particle filter, which is embedded in MVRPF, can move more particles into regions of high likelihood and hence can improve the tracking performance. Simulation results illustrate the effectiveness of the proposed method.