1-7hit |
Yusheng ZHANG Zhiheng ZHOU Bo LI Yu HUANG Junchu HUANG Zengqun CHEN
Person Re-Identification has received extensive study in the past few years and achieves impressive progress. Recent outstanding methods extract discriminative features by slicing feature maps of deep neural network into several stripes. Still there have improvement on feature fusion and metric learning strategy which can help promote slice-based methods. In this paper, we propose a novel framework that is end-to-end trainable, called Multi-level Slice-based Network (MSN), to capture features both in different levels and body parts. Our model consists of a dual-branch network architecture, one branch for global feature extraction and the other branch for local ones. Both branches process multi-level features using pyramid feature alike module. By concatenating the global and local features, distinctive features are exploited and properly compared. Also, our proposed method creatively introduces a triplet-center loss to elaborate combined loss function, which helps train the joint-learning network. By demonstrating the comprehensive experiments on the mainstream evaluation datasets including Market-1501, DukeMTMC, CUHK03, it indicates that our proposed model robustly achieves excellent performance and outperforms many of existing approaches. For example, on DukeMTMC dataset in single-query mode, we obtain a great result of Rank-1/mAP =85.9%(+1.0%)/74.2%(+4.7%).
Yu HUANG Zhiheng ZHOU Tianlei WANG Qian CAO Junchu HUANG Zirong CHEN
Vehicle detection is challenging in natural traffic scenes because there exist a lot of occlusion. Because of occlusion, detector's training strategy may lead to mismatch between features and labels. As a result, some predicted bounding boxes may shift to surrounding vehicles and lead to lower confidences. These bounding boxes will lead to lower AP value. In this letter, we propose a new approach to address this problem. We calculate the center of visible part of current vehicle based on road information. Then a variable-radius Gaussian weight based method is applied to reweight each anchor box in loss function based on the center of visible part in training time of SSD. The reweighted method has ability to predict higher confidences and more accurate bounding boxes. Besides, the model also has high speed and can be trained end-to-end. Experimental results show that our proposed method outperforms some competitive methods in terms of speed and accuracy.
Guoqi LIU Zhiheng ZHOU Shengli XIE Dongcheng WU
Vector field convolution (VFC) provides a successful external force for an active contour model. However, it fails to extract the complex geometries, especially the deep concavity when the initial contour is set outside the object or the concave region. In this letter, dynamically constrained vector field convolution (DCVFC) external force is proposed to solve this problem. In DCVFC, the indicator function with respect to the evolving contour is introduced to restrain the correlation of external forces generated by different edges, and the forces dynamically generated by complex concave edges gradually make the contour move to the object. On the other hand, traditional vector field, a component of the proposed DCVFC, makes the evolving contour stop at the object boundary. The connections between VFC and DCVFC are also analyzed. DCVFC maintains desirable properties of VFC, such as robustness to initialization. Experimental results demonstrate that DCVFC snake provides a much better segmentation than VFC snake.
Peijiang KUANG Zhiheng ZHOU Dongcheng WU
Recently, object-proposal methods have attracted more and more attention of scholars and researchers for its utility in avoiding exhaustive sliding window search in an image. Object-proposal method is inspired by a concept that objects share a common feature. There exist many object-proposal methods which are either in segmentation fashion or engineering categories depending on low-level feature. Among those object-proposal methods, Edge Boxes, which is based on the number of contours that a bounding box wholly contains, has the state of art performance. Since Edge Boxes sometimes misses proposing some obvious objects in some images, we propose an appropriate version of it based on our two observations. We call the appropriate version as Improved Edge Boxes. The first of our observations is that objects have a property which can help us distinguish them from the background. It is called object saliency. An appropriate way we employ to calculate object saliency can help to retrieve some objects. The second of our observations is that objects ‘prefer’ to appear at the center part of images. For this reason, a bounding box that appears at the center part of the image is likely to contain an object. These two observations are going to help us retrieve more objects while promoting the recall performance. Finally, our results show that given just 5000 proposals we achieve over 89% object recall but 87% in Edge Boxes at the challenging overlap threshold of 0.7. Further, we compare our approach to some state-of-the-art approaches to show that our results are more accurate and faster than those approaches. In the end, some comparative pictures are shown to indicate intuitively that our approach can find more objects and more accurate objects than Edge Boxes.
Jinjie LIANG Zhenyu LIU Zhiheng ZHOU Yan XU
Federated learning is a promising strategy for indoor localization that can reduce the labor cost of constructing a fingerprint dataset in a distributed training manner without privacy disclosure. However, the traffic generated during the whole training process of federated learning is a burden on the up-and-down link, which leads to huge energy consumption for mobile devices. Moreover, the non-independent and identically distributed (Non-IID) problem impairs the global localization performance during the federated learning. This paper proposes a communication-efficient FedAvg method for federated indoor localization which is improved by the layerwise asynchronous aggregation strategy and layerwise swapping training strategy. Energy efficiency can be improved by performing asynchronous aggregation between the model layers to reduce the traffic cost in the training process. Moreover, the impact of the Non-IID problem on the localization performance can be mitigated by performing swapping training on the deep layers. Extensive experimental results show that the proposed methods reduce communication traffic and improve energy efficiency significantly while mitigating the impact of the Non-IID problem on the precision of localization.
Zhiheng ZHOU Liang ZHOU Shengqiang LI
In wireless networks, how to provide reliable data transfer is an important and challenging issue due to channel fading and interference. Several approaches, e.g., Automatic Repeat reQuest (ARQ), Hybrid ARQ (HARQ) and Network Coding (NC), are used to enhance reliability of transmission in wireless networks. However, we note that these schemes implement the data recovery process for mixed unicast and multicast (MUM) communications by simply separating the process into two phases, unicast and multicast phase. This is inefficient and expensive. In this paper, we propose an efficient retransmission scheme with network coding for MUM transmission, aiming at improving bandwidth utilization. UMNC searches for coding opportunities from both unicast and multicast flows, which offer the potential benefit of improved recovery in the event of packet loss. We theoretically prove that UMNC can effectively reduce the total number of retransmissions and thus improve bandwidth efficiency, compared with existing schemes.
Ming DAI Zhiheng ZHOU Tianlei WANG Yongfan GUO
In many real application scenarios of image segmentation problems involving limited and low-quality data, employing prior information can significantly improve the segmentation result. For example, the shape of the object is a kind of common prior information. In this paper, we introduced a new kind of prior information, which is named by prior distribution. On the basis of nonparametric statistical active contour model, we proposed a novel distribution prior model. Unlike traditional shape prior model, our model is not sensitive to the shapes of object boundary. Using the intensity distribution of objects and backgrounds as prior information can simplify the process of establishing and solving the model. The idea of constructing our energy function is as follows. During the contour curve convergence, while maximizing distribution difference between the inside and outside of the active contour, the distribution difference between the inside/outside of contour and the prior object/background is minimized. We present experimental results on a variety of synthetic and natural images. Experimental results demonstrate the potential of the proposed method that with the information of prior distribution, the segmentation effect and speed can be both improved efficaciously.