Learning deep features from body and parts for person re-identification in camera networks

Abstract

In this paper, we propose to learn deep features from body and parts (DFBP) in camera networks which combine the advantages of part-based and body-based features. Specifically, we utilize subregion pairs to train the part-based feature learning model and predict whether they belong to positive subregion pairs. Meanwhile, we utilize holistic pedestrian images to train body-based feature learning model and predict the identities of the input images. In order to further improve the discrimination of features, we concatenate the part-based and body-based features to form the final pedestrian representation. We evaluate the proposed DFBP on two large-scale databases, i.e., Market1501 database and CUHK03 database. The results demonstrate that the proposed DFBP outperforms the state-of-the-art methods.

1 Introduction

The camera networks, as a kind of wireless sensor networks, have received considerable attention due to the potential value for the practical applications [1–4]. One of the most important applications of camera networks is person re-identification (re-ID) which is an issue of searching the same person from one camera sensor across different camera sensors with a probe image. Although substantial works have been made in this field, it remains some unsolved problems. This is because the appearances of pedestrians from different cameras are easily affected by many environmental factors, such as views, body pose, and lighting.

The existing approaches focus on two fundamental problems, i.e., feature representation and metric learning. On one hand, a lot of efforts expect to develop discriminative feature representations that are robust to the changes in views, pose, and illumination. Gray and Tao [5] proposed the Ensemble of Localized Features (ELF), each of which consisting of a feature channel, location, and binning information, to overcome the viewpoint changes. Bazzani et al. [6] presented the Symmetry-Driven Accumulation of Local Features (SDALFs) which considers symmetry and asymmetry perceptual principles to handle the environmental variances. Liao et al. [7] introduced the Local Maximal Occurrence (LOMO) features to analyze the horizontal occurrence of local features and maximize the occurrence in order to obtain a stable representation. On the other hand, some methods learn effective metrics. For example, Zheng et al. [8] proposed the Probabilistic Relative Distance Comparison (PRDC) to maximize the matching probability of positive pairs so that the distance between the positive pair is smaller than that of negative pair. Koestinger et al. [9] proposed the KISSME to learn an effective metric using equivalence constraints and had a good effect on the generalization performance. Liao et al. [7] introduced the Cross-view Quadratic Discriminant Analysis (XQDA) to simultaneously learn a discriminant low-dimensional subspace and a metric function on the derived subspace.

Recently, the convolutional neural network (CNN) is applied to person re-ID and achieves attractive performance on the large-scale databases. There are mainly two kinds of CNN models, i.e., verification models and identification models. The verification models [10–12] take pairs of images as the input and determine whether they belong to the same person or not, and the structure is shown in Fig. 1a. The verification models treat the person re-ID as a binary-class classification task and map two images of the same person into the nearby points; otherwise, the points are far apart. However, the verification models are not explicitly utilize re-ID labels. To make full use of the re-ID labels, the identification models [13, 14], which treat the person re-ID as a multi-class recognition task, are proposed, and the structure is shown in Fig. 1b. The identification models usually take holistic images as the input and determine whether images belong to their own labels. The identification models have shown great potentials on current large-scale person re-ID databases. Xiao et al. [15] jointly trained a identification model using multiple datasets and proposed a new dropout function to deal with a large number of classes. Furthermore, Zheng et al. [16] combined the verification and identification models by jointly learning three loss functions.

Fig. 1

The structures of a verification models and b identification models. a Verification models take pairs of pedestrian images as the input and predict whether they are the same person or not. b Identification models take pedestrian images as the input and predict their identity

The above methods utilize the holistic pedestrian images as the input and learn the body-based features, which discards the local characteristics of pedestrians. To overcome this drawback, Yi et al. [17] splited a pedestrian image into three horizontal parts and trained three part-CNNs. Cheng et al. [18] used the multi-channel CNN to learn local part features and modified the triplet loss function that requires the intra-class feature distances to be less than the inter-class ones. These existing part-based CNN methods learn the local features which are gradually ignored with the increase of neural network layers.

In this paper, we propose to learn deep features from body and parts (DFBP), which combines the advantages of part-based and body-based features. In the learning process, we simultaneously utilize two kinds of models, i.e., verification models and identification models, to learn and extract features. For learning part-based features, we split each pedestrian image into several horizontal subregions. We define the corresponding subregions from the same person as the positive subregion pairs and corresponding subregions from different persons as the negative subregion pairs. We train the verification models using the positive and negative subregion pairs and extract the subregion features from fully connected layer. Then, we utilize the weighted adding to aggregate all the subregion features from one pedestrian image and obtain the part-based feature. Meanwhile, we utilize the holistic images to train an identification model for the body-based features. Since we utilize two kinds of CNNs where they are also feed into different kinds of training samples, the learned part-based and body-based features are complementary. In order to further improve the discrimination of features, we concatenate the part-based and body-based features to form the final pedestrian representation.

The rest of this paper is organized as follows. In Section 2, we introduce the proposed method and the implementation process in detail. In Section 3, we verify the proposed DFBP on two large-scale databases, i.e., Market 1501 [19] and CUHK03 [20], and the experimental results demonstrate that the proposed DFBP outperforms the state-of-the-art methods. Finally, we conclude this paper in Section 4.

2 Method

The proposed DFBP is to learn powerful features by combining the part-based and body-based features in camera networks. The flowchart of the proposed DFBP is shown in Fig. 2. In this section, we introduce our model from three aspects, i.e., the part-based feature learning model, the body-based feature learning model, and the final representation for pedestrian images.

Fig. 2

The structures of a part-based and b body-based feature learning models

2.1 Part-based feature learning model

Part-based features could provide the effective local information which plays an important role in person re-ID. In order to learn part-based features, we apply the verification model and split each image into several horizontal subregions. We define the subregion pairs from the same person as positive subregion pairs as shown in Fig. 3a, and the subregion pairs from different person as negative subregion pairs as shown in Fig. 3b.

Fig. 3

Illustration of the a positive and b negative subregion pairs

There are two CNN models in the part-based feature learning model as shown in Fig. 2a. The two networks share the same weights which can map subregion pairs into the same feature subspace. Here, we utilize the widely used ResNet-50 [21] pretrained on ImageNet [22] as the CNN model. In the training stage, we resize all the subregions divided from pedestrian images into 256×256 and subtract the mean values computed from all the training subregions. Then, each subregion is randomly cropped to 224×224. Given subregion pairs as the input, we obtain two feature vectors f₁ and f₂ which are generated by average pooling at the end of ResNet-50 models as shown in Fig. 2a. In order to connect the two networks, we employ a non-parametric layer called square layer [16] to compare the features f₁,f₂. The input of square layer is f₁ and f₂, and the output is formulated as:

$$ {f_{s}}=(f_{1}-f_{2})^{2} $$

(1)

where f₁,f₂, and f _s are the 2048 dimensional vectors for the ResNet-50 model.

After this layer, we utilize the dropout strategy [23], which introduces zero values at special locations, to regularize the model. Then, we utilize a convolutional layer to turn the resulting vector f _s to a 2-dimensional vector (p₁,p₂) which represents the prediction probability of the input subregion pair belonging to the same person. The convolutional layer takes f _s as the input and has two kernels with the size of 1×1×2048. We do not add Rectified Linear Unit (ReLU) after this layer and utilize the softmax to predict the probability:

$$ {y_{i}}=\frac{e^{a_{i}}}{\sum_{k=1}^{2}{e^{a_{k}}}}, \forall~{i}~\in~{1,2} $$

(2)

where a _i is the activation value of the i-th neuron in the last convolutional layer.

We utilize the cross-entropy as the cost function of part-based feature learning model which is formulated as:

$$ {L_{V}}=-\sum_{i=1}^{2}{p_{i}\ \text{log}\ {y_{i}}} $$

(3)

where y _i ∈ [0,1] is the prediction probability and p _i is the true probability. If the input is the positive subregion pair, then p₁=1 and p₂=0; otherwise, p₁=0 and p₂=1.

Since the two CNN models share the same weights, we can extract subregion features from any ResNet-50 model. We treat f₁ or f₂ as the feature for each subregion.

2.2 Body-based feature learning model

Body-based features could provide holistic information of pedestrian, and they are complementary to part-based features. We adopt the identification model for learning body-based features as shown in Fig. 2b and also utilize the ResNet-50 [21] as the CNN model. We take the holistic pedestrian images as the input of ResNet-50 model and utilize the same preprocessing for each pedestrian image as mentioned in Section 2.1. Then, we obtain the feature f$\in \mathbb {R}\ ^{2048*1}$ for each pedestrian image which are generated by the average pooling at the end of ResNet-50 model. Similarly, we employ the dropout strategy to prevent over-fitting. Then, we utilize the convolutional layer to turn the resulting vector f into a C dimensional vector where each element represents the prediction probability of the input pedestrian image belonging to one label. The class number of pedestrian is C, so the convolutional layer has C kernels with the size of 1×1×2048. For example, since the Market1501 database has 751 pedestrian classes, C is set to 751. We do not add ReLU after this layer. We also use the softmax unit to normalize the output and obtain the prediction probability. It is formulated as:

$$ {y_{j}}=\frac{e^{a_{j}}}{\sum_{k=1}^{C}{e^{a_{k}}}}, \forall~{j}~\in~{1,2,\ldots,C} $$

(4)

where a _j is the activation value of the j-th neuron in the last convolutional layer.

We utilize the cross-entropy loss for body-based feature learning model:

$$ {L_{I}}=-\sum_{j=1}^{K}{q_{j}\ \text{log}\ {y_{j}}} $$

(5)

where y _j ∈ [0,1] is the prediction probability of the input belonging to the j-th class and q _j is the true probability. If n is the true class, then q _n =1; otherwise, q _j =0 for all j.

After training the body-based feature learning model, we extract f as the body-based feature for each pedestrian image.

2.3 Final representation

In order to obtain a powerful feature representation, we combine the part-based features and the body-based features to form DFBP. The flowchart of the combination strategy is shown in Fig. 4.

Fig. 4

The flowchart of the proposed DFBP extraction

We split each pedestrian image into several subregions, and the number of subregions is the same as mentioned in Section 2.1. Then, we extract the feature for each subregion and aggregate them by weighted adding to obtain the part-based feature:

$$ P=\alpha_{1}{p_{1}}+\alpha_{2}{p_{2}}+\cdots+\alpha_{n}{p_{n}} $$

(6)

where p _i (i = 1, 2, …, n) represents the feature of the i-th subregion in a pedestrian image, α _i (i = 1, 2, …, n) is the corresponding coefficient to control the weight of the i-th subregion, and n is the total number of subregions in a pedestrian image.

Finally, we concatenate the part-based feature and body-based feature to form the final DFBP representation:

$$ F=[\eta P,\varphi B] $$

(7)

where [.,.] indicates directly concatenating the two vectors, B represents the body-based feature, η and φ are the corresponding coefficients to control the weights of P and B respectively, and F$\in \mathbb {R}\ ^{4096*1}$ represents the final feature DFBP. We simply use η = 1 and φ = 1. This may not be the best choice for a particular database. But if we do not foreknow the distribution of the database, it is a simple and straight-forward choice.

3 Experiments

In this section, we verify the effectiveness of the proposed DFBP on two large-scale databases. We first introduce the experimental setup in Section 3.1 and then show the results on the Market1501 database [19] and the CUHK03 database [20] in Sections 3.2 and 3.3 respectively.

3.1 Experimental setup

We implement the proposed model using the MatConvNet package [24]. In the training phase, in order to fine-tune the pre-trained ResNet-50 [21], we modify the last convolution layer to have 751 kernels for the Market1501 database and 1367 kernels for the CUHK03 database. We set the maximum number of training epochs to 75. The learning rate is initialized as 0.1 and decreased to 0.01 after 70 epochs. We utilize stochastic gradient descent (SGD) to update the parameters and set the weight decay to 0.0005. The batch size for training part-based and body-based feature models is 24. In the dropout layer, we set the dropout rate to 0.5. For the part-based feature learning model, we split each pedestrian image into three horizontal subregions and empirically set α₁ = 0.5, α₂ = 0.3, and α₃ = 0.2. As the input of part-based feature learning model, the ratio between positive subregion pairs and nagative subregion pairs is 1:1.

3.2 Market1501 database

Market1501 database [19] is a newly released large-scale person re-ID database which is collected in a university campus. The database contains 32,669 annotated bounding boxes of 1501 identities. According to the database setting, it includes 12,936 training images of 751 identities and 19,732 test images of 750 identities. In addition, there are 3368 query images about the same 750 identities as the test set. Images of each identity are captured by at most six manually set cameras, and each person has 3.6 images on average at each viewpoint. All images are automatically detected by the Deformable Part based Model (DPM) [25]. So the database is close to realistic setting.

In the experiments, we select a pedestrian from the query set and retrieve the truth images in the test set. We utilize the mean average precision (mAP) and rank-1 accuracy to evaluate the performance. Here, rank-i accuracy denotes the probability whether one or more correctly matched images appear in top-i. The evaluation results of the single query setting and multiple query setting are shown in Table 1. From the results, we can see that the proposed DFBP achieves the highest classification accuracy of 81.71% (rank-1) and 60.86% (mAP) in the single query setting, and 87.01% (rank-1) and 72.21% (mAP) in the multiple query setting. The performance of the proposed DFBP is better than the part-based features and body-based features, because we make full use of the complementarity between the part-based features and body-based features. The proposed DFBP and Verif.-Identif [16] utilize the verification and identification models. Noticeably, the proposed DFBP outperforms the Verif.-Identif in all situations due to considering the structural information of pedestrian by learning part-based features.

Table 1 The rank-1 precious (%) and mAP (%) on the Market-1501 database

We further evaluate the performance of the proposed DFBP between all camera pairs, and the results are shown in Fig. 5. The two matrices are confusion matrices in the single query setting. In the two matrices, the accuracy q _ij (1 ≤i ≤ 6, 1 ≤j ≤ 6) indicates the pedestrian in the i-th camera as the query image and the pedestrian in the j-th camera as the gallery image with the same labels. The camera 6 is a 720 ×576 low-resolution camera and captures distinct background with the other high-resolution cameras. From the results, we can see that the re-ID accuracies between camera 6 and the others are relatively high. We also compute the average cross-camera accuracy to obtain 50.17% mAP and 56.93% rank-1 accuracy, respectively. The average cross-camera indicates the average of all accuracies in the matrix except the accuracies on the diagonal of the matrix in Fig. 5. The proposed DFBP largely improves the performance compared with the previous reported results, such as 10.51% mAP, 13.72% rank-1 accuracy in [19] and 48.42% mAP, 54.42% rank-1 accuracy in [16]. These results indicate that the proposed DFBP could also learn the discriminative features under different viewpoints.

Fig. 5

Re-ID performance between camera pairs on Market1501 database. a Rank-1 accuracy. b mAP accuracy

3.3 CUHK03 database

CUHK03 database [20] includes 14,097 images of 1467 identities which is captured by six cameras over months in the Chinese University of Hong Kong (CUHK) campus. Each pedestrian is observed by two disjoint camera views and has an average of 4.8 images in each view. The database provides two kinds of bounding boxes, i.e., manually annotated bounding boxes and bounding boxes detected by DPM [25]. Since DPM is closer to the realistic setting, we evaluate the proposed DFBP on the bounding box detected by DMP. We separate the database into training data including 1367 pedestrians and test data including 100 pedestrians. This partitioning is independently implemented 10 times, and the average is reported. We evaluate the proposed DFBP by the single-query setting and multiple-query setting.

We utilize the rank-i and mAP to evaluate the proposed DFBP, and the results in the single-shot setting are shown in Table 2. There is only a correct image in the gallery set from the other camera. From the Table 2, we can see that the proposed DFBP achieves the highest classification accuracy of 75.9% rank-1 accuracy, 93.8% rank-5 accuracy, 97.9% rank-10 accuracy, and 79.9% mAP. This demonstrates the effectiveness of the proposed DFBP once again.

Table 2 Comparison with the state-of-the-art methods on the CUHK03 dataset under the single-shot setting

In addition, we evaluate the proposed DFBP in the multi-shot setting as shown in Table 3. We use all the images from the other camera as gallery images, and the number of the images is about 500. From Table 3, we can see that the proposed DFBP also achieves the highest accuracy owing to combining the part-based features and body-based features. Furthermore, we represent some re-ID samples in Fig. 6. The first column shows query images and the remaining columns are retrieved images which are sorted according to the similarity scores from left to right. Most retrieved images are correct. Although there are several incorrect images, such as the last eight images in the sixth row, we think it is a reasonable prediction because the pedestrians are similar to the query in appearance and color.

Fig. 6

Samples of pedestrian retrieval on CUHK03 database in the multi-shot setting. The first column indicates query images, and remaining columns are gallery images which are sorted according to the scores from left to right

Table 3 Comparison with the state-of-the-art methods on the CUHK03 dataset using the multi-shot setting

4 Conclusions

In this paper, we have proposed a novel feature learning method named DFBP in sensor networks, which simultaneously considers the part-based features and body-based features. We utilize subregion pairs and holistic pedestrian images to train part-based and body-based feature learning models respectively. By integrating the two kinds of features, we obtain a discriminative feature. We have proved the effectiveness of the proposed DFBP on two databases which are realistic and challenging, and the proposed DFBP have achieved better performance than the state-of-the-art methods.