基于GAN的无监督域自适应行人重识别

针对无监督域自适应行人重识别中存在的聚类不准确导致网络识别准确率低的问题,提出一种基于生成对抗网络的无监督域自适应行人重识别方法。首先通过在池化层后使用批量归一化层、删除一层全连接层和使用Adam优化器等方法优化CNN模型;然后基于最小错误率贝叶斯决策理论分析聚类错误率和选择聚类关键参数;最后利用生成对抗网络调整聚类,有效提升了无监督域自适应行人重识别的识别准确率。在源域Market-1501和目标域DukeMTMC-reID下进行实验,mAP和Rank-1分别达到了53.7%和71.6%。     

基于FPGA深度学习的行人重识别研究

文章提出了一种基于深度学习的行人重识别算法,并利用FPGA硬件对该算法进行实验,最终实现一种新型的行人重识别技术的研究。     

联合特征细化和耐噪声对比学习的无监督行人重识别

针对无监督行人重识别(person re-identification, ReID)中行人特征表达不充分以及聚类过程产生噪声标签的问题,提出一种联合特征细化和耐噪声对比学习的无监督ReID方法。首先,为丰富无标记的行人表征,设计了非局部通道细化模块(non-local channel refinement module, NCRM)对关键特征信息进行加权强化,其融合了非局部通道的重要特征来捕获无标记数据的类间区别表征,形成更具有鉴别度的特征描述符。其次,考虑到特征的充分表达,采用广义均值(generalized mean, GEM)池化自适应调整参数来增强不同细粒度区域信息的提取能力。再次,为了减轻噪声标签对网络的负面影响,设计了耐噪声的动态对比均衡(dynamic contrastive equilibrium, DCE)损失函数进行无监督联合学习。最终,在两个公共数据集上的实验结果验证了所提方法的有效性和先进性,mAP分别达到了83.1%和71.9%,优于其他先进方法。     

基于多尺度批量特征丢弃网络的行人重识别研究

针对行人重识别存在的遮挡和姿势变化问题和目前网络识别率低的缺陷,提出了不同空间维度的多分支行人重识别网络模型。首先利用IBN-Net50-a为基础骨干网络提取特征;然后对最后两层卷积层融合批量特征丢弃方法,以增强局部区域专注特征学习;最后拼接不同维度的特征,获得更多浅层、深层的有用信息。在网络训练时,采用三元组损失和标签平滑损失联合策略训练。使用三个常用的基准数据集Market1501、DukeMTMC-reID、CUHK03进行实验验证,并按照主流策略划分数据集。实验结果表明,所提方法的特征泛化能力较好,其中在Market1501数据集的Rank-1和平均准确率（mAP）分别达到95.3%和86.8%;在DukeMTMC-reID数据集的Rank-1和mAP分别达到88.5%和75.9%;在CUHK03数据集的Rank-1和mAP分别达到80.9%和77.8%。     

基于局部特征相关性的行人重识别

针对因行人重识别中行人图像的背景、姿势等差异和行人图像的属性相似导致的行人重识别准确率低的问题,提出了一种基于特征相关性学习的行人重识别方法,加强了局部特征提取的准确性和相关性,充分考虑了行人身体部位和其他部位之间的相关性。首先使用Resnet-50网络和人体关键点估计模型来提取局部特征和全局特征,然后对局部特征和全局特征进行分支操作,对局部特征进行相关性学习,加强各个局部特征与其他局部特征之间的联系,提高相似属性行人图像之间的鉴别性;对全局特征通过池化操作提取更为精确的全局特征,最后将相关性学习后的局部特征和池化后的全局特征进行拼接作为预测行人身份的对比特征。此模型在Market-1501数据集及DukeMTMC-ReID数据集上的mAP指标分别达到了85.6%和76.1%,验证了所提网络模型的有效性。     

基于PCB特征加权的行人重识别算法

为充分挖掘行人重识别特征,最近流行的PCB算法给出了一种特征均匀分块并通过RPP网络对齐特征的方法。PCB算法充分发挥了局部特征的作用,有效提高了行人重识别的准确率。为进一步提高行人重识别的性能,本文基于全局特征与局部特征对网络性能的影响差异提出了一种特征加权的PCB行人重识别算法。在典型的行人识别数据库Market1501、DukeMTMC-Reid上的实验结果表明:所提算法具有更好的首中准确率（Rank1）和平均准确率（mAP）;相比与经典的PCB+RPP算法,所提算法在Market1501数据集上Rank1提高了0.8%,mAP提高了4.5%;在DukeMTMC-Reid数据集上Rank1提高了5.5%,mAP提高了约7%。      

基于卷积-时空特征融合的行人重识别

考虑到目前监控系统中在地域上的局限性,提出一种利用监控系统中相机的时空分布特征与图像的卷积特征融合的方法.在训练时除了需要训练用于提取图像特征的卷积神经网络,还需要计算出相机之间的时空分布.在预测时,先利用图像特征计算出检测图像的得分,然后通过对数加权的方法与时空特征融合,从而获得最终排序得分.结果表明,模型具有较好的...     

MTNet: Mutual tri-training network for unsupervised domain adaptation on person re-identification

The existing unsupervised domain adaptation (UDA) methods on person re-identification (re-ID) often employ clustering to assign pseudo labels for unlabeled target domain samples. However, it is difficult to give accurate pseudo labels to unlabeled samples in the clustering process. To solve this problem, we propose a novel mutual tri-training network, termed MTNet, for UDA person re-ID. The MTNet method can avoid noisy labels and enhance the complementarity of multiple branches by collaboratively training the three different branch networks. Specifically, the high-confidence pseudo labels are used to update each network branch according to the joint decisions of the other two branches. Moreover, inspired by self-paced learning, we employ a sample filtering scheme to feed unlabeled samples into the network from easy to hard, so as to avoid trapping in the local optimal solution. Extensive experiments show that the proposed method can achieve competitive performance compared with the state-of-the-art person re-ID methods.     

Modeling feature distances by orientation driven classifiers for person re-identification

To tackle the re-identification challenges existing methods propose to directly match image features or to learn the transformation of features that undergoes between two cameras. Other methods learn optimal similarity measures. However, the performance of all these methods are strongly dependent from the person pose and orientation. We focus on this aspect and introduce three main contributions to the field: (i) to propose a method to extract multiple frames of the same person with different orientations in order to capture the complete person appearance; (ii) to learn the pairwise feature dissimilarities space (PFDS) formed by the subspaces of similar and different image pair orientations; and (iii) within each subspace, a classifier is trained to capture the multi-modal inter-camera transformation of pairwise image dissimilarities and to discriminate between positive and negative pairs. The experiments show the superior performance of the proposed approach with respect to state-of-the-art methods using two publicly available benchmark datasets.     

Comprehensive feature fusion mechanism for video-based person re-identification via significance-aware attention

Video-based person re-identification (Re-ID) is of important capability for artificial intelligence and human–computer interaction. The spatial and temporal features play indispensable roles to comprehensively represent the person sequences. In this paper, we propose a comprehensive feature fusion mechanism (CFFM) for video-based Re-ID. We use multiple significance-aware attention to learn attention-based spatial–temporal feature fusion to better represent the person sequences. Specifically, CFFM consists of spatial attention, periodic attention, significance attention and residual learning. The spatial attention and periodic attention aim to respectively make the system focus on more useful spatial feature extracted by CNN and temporal feature extracted by the recurrent networks. The significance attention is to measure the two features that contribute to the sequence representation. Then the residual learning plays between the spatial and temporal features based on the significance scores for final significance-aware feature fusion. We apply our approach to different representative state-of-the-art networks, proposing several improved networks for improving the video-based Re-ID task. We conduct extensive experimental results on the widely utilized datasets, PRID-2011, i-LIDS-VID and MARS, for the video-based Re-ID task. Results show that the improved networks perform favorably against existing approaches, demonstrating the effectiveness of our proposed CFFM for comprehensive feature fusion. Furthermore, we compare the performance of different modules in CFFM, investigating the varied significance of the different networks, features and sequential feature aggregation modes.     

HCFN: Hierarchical cross-modal shared feature network for visible-infrared person re-identification

Compared with traditional visible–visible person re-identification, the modality discrepancy between visible and infrared images makes person re-identification more challenging. Existing methods rely on learning efficient transformation mechanisms in paired images to reduce the modality gap, which inevitably introduces noise. To get rid of these limitations, we propose a Hierarchical Cross-modal shared Feature Network (HCFN) to mine modality-shared and modality-specific information. Since infrared images lack color and other information, we construct an Intra-modal Feature Extraction Module (IFEM) to learn the content information and reduce the difference between visible and infrared images. In order to reduce the heterogeneous division, we apply a Cross-modal Graph Interaction Module (CGIM) to align and narrow the set-level distance of the inter-modal images. By jointly learning two modules, our method can achieve 66.44% Rank-1 on SYSU-MM01 dataset and 74.81% Rank-1 on RegDB datasets, respectively, which is superior compared with the state-of-the-art methods. In addition, ablation experiments demonstrate that HCFN is at least 4.9% better than the baseline network.     

Camera style transformation with preserved self-similarity and domain-dissimilarity in unsupervised person re-identification

The inconsistency caused by different factors, such as different camera imaging methods, complex imaging environments, and changes in light, present a huge challenge to person re-identification (re-ID). Unsupervised domain adaptation (UDA) can solve the inconsistency issue to a certain extent, but different datasets may not have any overlapping of people’s identities. Therefore, it is necessary to pay attention to people’s identities in solving domain-dissimilarity. A camera imaging style transformation with preserved self-similarity and domain-dissimilarity (CSPSD) is proposed to solve the cross-domain issue in person re-ID. First, CycleGAN is applied to determine the style conversion between source and target domains. Intra-domain identity constraints are used to maintain identity consistency between source and target domains during the image style transformation process. Maximum mean difference (MMD) is used to reduce the difference in feature distribution between source and target domains. Then, a one-to-n mapping method is proposed to achieve the mapping between positive pairs and distinguish negative pairs. Any sample image from the source domain and its transformed image or a transformed image with the same identity information compose a positive pair. The transformed image and any image from the target domain compose a negative pair. Next, a circle loss function is used to improve the learning speed of positive and negative pairs. Finally, the proposed CSPSD that can effectively reduce the difference between domains and an existing feature learning network work together to learn a person re-ID model. The proposed method is applied to three public datasets, Market-1501, DukeMTMC-reID, and MSMT17. The comparative experimental results confirm the proposed method can achieve highly competitive recognition accuracy in person re-ID.     

Grafted network for person re-identification

Convolutional neural networks have shown outstanding effectiveness in person re-identification (re-ID). However, the models always have large number of parameters and much computation for mobile application. In order to relieve this problem, we propose a novel grafted network (GraftedNet), which is designed by grafting a high-accuracy rootstock and a light-weighted scion. The rootstock is based on the former parts of ResNet-50 to provide a strong baseline, while the scion is a new designed module, composed of the latter parts of SqueezeNet, to compress the parameters. To extract more discriminative feature representation, a joint multi-level and part-based feature is proposed. In addition, to train GraftedNet efficiently, we propose an accompanying learning method, by adding an accompanying branch to train the model in training and removing it in testing for saving parameters and computation. On three public person re-ID benchmarks (Market1501, DukeMTMC-reID and CUHK03), the effectiveness of GraftedNet is evaluated and its components are analyzed. Experimental results show that the proposed GraftedNet achieves 93.02%, 85.3% and 76.2% in Rank-1 and 81.6%, 74.7% and 71.6% in mAP, with only 4.6M parameters.     

Interactive information module for person re-identification

In person re-identification (Re-ID) task, multi-branch networks acquire better performance by combining global features and local features. Obviously, local branch can obtain detailed information of person pictures but may work on invalid regions when person pictures have imprecise bounding boxes. On the contrary, global branch can be aware of the position of person but hard to acquire detailed information of person pictures. Meanwhile, lots of multi-branch networks ignore mutual information among different branches. Therefore, it is necessary to enhance interaction of global branch and local branch. For this purpose, we propose Interactive Information Module (IIM). IIM includes two components named Global-map Attention Module (GAM) and Labeled-class Mutual Learning (LML), respectively. GAM leverages heatmaps generated by global branch to guide calculation of local attention and obtains a composite global feature by combining local features. GAM relys more on the performance of global branch which decides the quality of heatmaps. To improve performance of global branch, we propose LML to promote convergent rate of global branch. Extensive experiments implemented on Market-1501, DukeMTMC-ReID, and CUHK03-NP datasets confirm that our method achieves state-of-the-art results.