Most existing re-ID methods only take identity labels of pedestrians into consideration.
However, we ﬁnd the attributes, containing detailed local descriptions, are beneﬁcial in allowing the re-ID model to learn more discriminative feature representations.
属性包含了详细的局部描述。 属性有助于re-ID 模型去学习更有辨别的特征表达
in this paper, based on the complementarity of attribute labels and ID labels, we propose an attribute-person recognition (APR) network, a multi-task network which learns a re-ID embedding and at the same time predicts pedestrian attributes.
基于 属性 标签和 身边（ID）标签的互补性，提出了——
attribute-person recognition (APR) network (属性-人物识别网络)
We manually annotate attribute labels for two large-scale re-ID datasets, and systematically investigate how person re-ID and attribute recognition beneﬁt from each other. In addition, we re-weight the attribute predictions considering the dependencies and correlations among the attributes.
- 给两个大规模re-ID数据集 手动标注了属性标签
- 系统地调查了人物 的re-ID 和属性识别如何互利 考虑到属性之间的依赖性和相关项，re-weight（重新加权）了属性预测
The experimental results on two large-scale re-ID benchmarks demonstrate that by learning a more discriminative representation, APR achieves competitive re-ID performance compared with the state-of-the-art methods.
We use APR to speed up the retrieval process by ten times with a minor accuracy drop of 2.92% on Market-1501. Besides, we also apply APR on the attribute recognition task and demonstrate improvement over the baselines.
通过学习更有辨别力的表达，APR 拥有了和state-of-the-art方法相比较的竞争力。 我们使用APR将检索过程加速十倍，精度下降幅度为2.92％ 在属性识别任务上应用APR，并展示出了改进。
Attributes describe detail information for a person, including gender, accessory, the color of clothes, etc .
属性： 性别、配饰、衣服的颜色 etc.
In this paper, we aim to improve the performance of large-scale person re-ID, using complementary cues （互补线索）from attribute labels.
The motivation of this paper is that existing large-scale pedestrian datasets for re-ID contains only annotations of identity labels, we believe that attribute labels are complementary with identity labels in person re-ID.
现有数据集仅标注了身份信息，我们却坚信 属性标签在re-ID任务上 和 身份标签是互补的。
The effectiveness of attribute labels is three-fold:
First, training with attribute labels improves the discriminative ability of a re-ID model.
Attribute labels can depict pedestrian images with more detailed descriptions.
These local descriptions push pedestrians with similar appearances closer to each other and those different away from each other Second, detailed attribute labels explicitly guide the model to learn the person representation by designated human characteristics. With the attribute labels, the model is able to learn to classify the pedestrians by explicitly focusing on some local semantic descriptions, which greatly ease the training of models.
Third, attributes can be used to accelerate the retrieval process of re-ID The main idea is to ﬁlter out some gallery images that do not have the same attributes as the query.
In  , the PETA dataset is proposed which contains both attribute and identity attributes. However, PETA is comprised of small datasets and most of the datasets only contain one or two images for an identity.
When using attributes for re-ID, attributes can be used as auxiliary information for low level features  or used to better match images from two cameras [10–12] .
 属性被用作辅助信息 [10-12] 属性被用于更好地匹配来自两个摄像机的图像
In recent years, some deep learning methods are proposed [13–15] . In these works, the network is usually trained by several stages. Franco et al.
 propose a coarse-to-ﬁne learning framework. The network is comprised of a set of hybrid deep networks, and one of the networks is trained to classify the gender of a person. In this work, the networks are trained separately and thus may over- look the complementarity of the general ID information and the attribute details. Besides, since gender is the only attribute used in the work, the correlation between attributes is not leveraged in  .
- 其中一个网络 用于对人的性别进行分类。
In [14,15] , the network is ﬁrst trained on an independent attribute dataset, and then the learned information is transferred to the re-ID task.
A work closest to ours consists of  , in which the CNN embedding is only optimized by the attribute loss. We will show that by combining the identiﬁcation and attribute recognition with an attribute re-weighting module, the APR network is superior to the method proposed in  .
 CNN嵌入仅通过属性损失进行优化 VS 我们将通过将 带有属性重置权重模块的属性识别 和 身份识别 相结合来证明 ，APR网络优于中提出的方法
First, our work systematically investigates how person re-ID and attribute recognition beneﬁt each other by a jointly learned network.
通过联合学习网络，我们系统地调查了 re-ID 和属性识别是如何互利的
On the one hand, identity labels provide global descriptions for person images, which have been proved effective for learning a good person representation in many re-ID works [17-19] On the other hand, attribute labels provide detailed local descriptions.
身份标签 提供全局描述 属性标签提供详细局部描述 —— 由此实现更高准确率的 属性识别和 re-ID 识别。
Second, in previous works, the correlations of attributes are hardly considered.
In fact, many attributes usually cooccur for a person, and the correlations of attributes may be helpful to re-weight the prediction of each attribute. We thereby introduce an Attribute Re-weighting Module to utilize correlations among attributes and optimize attribute predictions.
In this paper, we propose the attribute-person recognition (APR) network to exploit both identity labels and attribute annotations for person re-ID.
By combining the attribute recognition task and identity classiﬁcation task, the APR network is capable of learning more discriminative feature representations for pedestrians, including global and local descriptions.
结合属性识别任务 ** 和 **身份分类任务 ， APR网络可以学习 更有辨别力的特征表达， 包括全局和局部描述。
Speciﬁcally, we take attribute predictions as additional cues for the identity classiﬁcation. Considering the dependencies among pedestrian attributes, we ﬁrst re-weight the attribute predictions and then build identiﬁcation upon these re-weighted attributes descriptions.
我们将属性预测 作为 身份分类的附加线索。 考虑属性之间的依赖性，我们首先 re-weight 了 属性预测 并且 在这些re-weight了的属性描述上 构建了身份。
The attribute is also used to speed up the retrieval process by ﬁltering out the gallery images with different attribute from the query image.
In the experiment, we show that by applying the attribute acceleration process, the evaluation time is saved to a signiﬁcant extent.
We evaluate the performance of the proposed method APR on two large-scale re-ID datasets and an attribute recognition dataset. The experimental results show that our method achieves competitive re-ID accuracy to the state-of-the-art methods.
APR的性能 和state-of-the-art methods 有得一比
In addition, we demonstrate that the proposed APR yields improvement in the attribute recognition task over the baseline in all the testing datasets. APR 在属性识别任务上有所改进。
(1) We have manually labeled a set of pedestrian attributes for the Market-1501 dataset and the DukeMTMC-reID dataset. Attribute annotations of both datasets are publicly available on our website ( https://vana77.github.io ).
(2) We propose a novel attribute-person recognition (APR) framework. It learns a discriminative Convolutional Neural Network (CNN) embedding for both person re-identiﬁcation and attributes recognition.
(3) We introduce the Attribute Re-weighting Module (ARM), which corrects predictions of attributes based on the learned dependency and correlation among attributes.
引入了 属性 re-weighing 模块（ARM）, 它根据学习到的属性之间的相关性和依赖性来纠正属性的预测。
(4) We propose an attribute acceleration process to speed up the retrieval process by ﬁltering out the gallery images with different attribute from the query image. The experiment shows that the size of the gallery is reduced by ten times, with only a slight accuracy drop of 2.92%. 提出 了属性加速过程。
(5) We achieve competitive accuracy compared with the state- of-the-art re-ID methods on two large-scale datasets, i.e., Market-1501  and DukeMTMC_reID  . We also demonstrate improvements in the attribute recognition task.
效果和state- of-the-art re-ID methods 有得一比。 在属性识别任务上有改进。