1 Introduction

Person re-identification refers a task of associating the same person in cross camera views. But there are still many unsolved problems that due to changes in view angle, bodies pose, illumination, occlusion and background clutter. To address these difficulties, researchers are actively working on design feature [11, 15, 26, 42, 43, 63] and learning a distance [9, 12, 29, 34, 41, 44, 62], or both jointly [1, 2, 13, 14]. For feature representation, several effective approaches have been proposed, for example, Local Maximal Occurrence (LOMO) [13], GOG [26], KCCA [15], Mid-level [63], HOG3D [44], and SCNCD [43]. These hand-crafted features make impressive improvements in person re-identifications research. However, all of these feature descriptors are based on global appearance, or only taking the individual itself into account, without considering the attachment that captured by camera. Naturally, the interference caused by these areas that covered by the new appendages is ignored.

To solve these problems, we propose a novel framework for person re-identification by splitting the person picture in horizontal and vertical directions. This inspiration comes from SCSP [2], which divided target into four sub-regions in horizontal directions. The first difference is that we add the vertical direction of the regional division. It is very effective when a camera is shot in the front and another camera in the side. Moreover, the major improvement is that we do not regard the results of all regions as valid values. By contrast, we take a comparison in the neighboring region, the region getting a lower score will be treated as invalid one. This is effective for removing interference of Damaged Region.

Before splitting the target, we enlarge it with directional to add the useful information by bilinear interpolation algorithm. As a matter of fact, this measure combine with regional validity judgment can reduce the impact of viewpoint variation.

The features we have used are HSV and SILTP, and the extract method is the same as LOMO [13], the different is that we used the local mean instead local maximal. The mean value can increase the anti-interference of noise and improve the robustness, and reduce the randomness that brought by the maximum.

Our contributions can be summarized as follows:

  • we present an efficient feature representation which combines HSV and SILTP after obtaining the mean value in the horizontal patches, and it is more robust than lomo [13].

  • We proposed a novel framework by dividing the image into sub-region both in horizontal and vertical directions and used the higher score one as the valid region after comparison. This operation is more conducive to re-identifying work than not dividing the image or splitting the image in a single direction.

  • We enlarge the image with direction before splitting them. The results show that the proposed model achieves an outstanding performance compared with the state-of-the-art person re-identification approaches.

The rest of this paper is organized as followed. Section 2, reviewing related works, and describing the framework of the proposed approach. Section 3 describe the details of our proposed method, include: how to extract the feature; the concept of Damaged Region (DR); the details of the image partition and directional enlargement. The experiments and results are in Section 4. We finally make a conclusion and discuss possible future works in Section 5.

2 Related work

Recently, Computer vision [19, 24, 38] has attracted wide attention,and image categorization, image cropping and segmentation and image recognition are important yet challenging tasks in this fields. Under the efforts of many researchers, many algorithms are proposed for image categorization [47, 49, 50, 55, 58, 60, 61]. More specifically, the [39, 40, 49, 55, 58] focus on the aerial image, and [47, 60, 61] belongs to fine-grained image categorization. In [50], Zhang proposed to learn object-shaped and directional receptive fields for image categorization. As one of the most basic photo manipulation processes, photo cropping [46, 56, 57] and image segmentation [48, 54, 59] attracted wide attention. In [17], the random forest model, which is useful for image segmentation and categorization, is constructed by semi-supervised node splitting. The image recognition mainly includes the action recognition [20, 21, 36, 64], activity recognition [16, 20, 21, 23, 25] and person re-identification. Semantic photo retargeting [51] and human fixations prediction [52] are the key technologies in action recognition. However, considering the limited storage space, we mainly discuss person re-identification in this paper.

Given a probe image, the person re-identification system aims at identifying a set of matching images from a gallery set, which is mostly captured by a different camera. It plays a crucial role in target tracking [4, 18, 27], target retrieval [6, 37] and predict work [22, 30]. In the past five years, person re-identification has attracted great attention and made great achievements with the joint efforts of the computer vision researchers. However, re-identifying results may be inaccurate and not robust due to appearance features variation caused by various viewpoint changes, occlusion, and individual movement factors. In addition, since people are mainly distinguishable by their clothing under surveillance setting, similar clothes among different persons add further challenges.

Recently, several research methods have been proposed for solving these problems in person re-identification. Design feature and distance metric are two key points. Many existing approaches try to build robust feature representations to describe a person’s appearance under various conditions [11, 15, 26, 42, 43, 63]. Liao et al. [13] proposed an efficient feature representation called Local Maximal Occurrence (LOMO), using color and Scale Invariant Local Ternary Pattern (SILTP) histograms to represent picture appearance in a high dimension. This method locally constructs a histogram of pixel features, and then takes its maximum values within horizontal strips to overcome viewpoint variations while maintaining local discrimination. And mid-level filters (Mid-Level) in [63] was learned from patch clusters with coherent appearance gained by pruning hierarchical clustering trees to obtain view-invariant and discriminative features. In [20], it provided a mid-level feature representation for activities to captures the intrinsic properties for recognition work. In [26] a method called Gaussian of Gaussian (GOG) descriptor is proposed, which based on a shierarchical Gaussian distribution of pixel features. Using a complex algorithm in Riemannian mathematics, resulting in a considerable amount of time to extract the features. Chen et al. [2] partition an image into four non-overlap horizontal stripe regions, and each stripe region can be described by four visual cues which are organized as HSV1/HOG, HSV2/SILPT, LAB1/SILPT, LAB2/HOG. The optimization function adopt Alternating Direction Method of Multipliers (ADMM), which needs to spend more time to learn the coefficient matrix than the general metric learning methods. Besides, it is common to use multiple features [45, 53] to represent an object, and these features usually high-order correlated. Specifically, in the paper [45] a new multimodal feature integration framework was proposed, which construct a feature correlation hypergraph (FCH) to model the high-order relations among multimodal features.

Learning a robust distance metric is another important part in re-identification, which formulates the person re-identification as a distance metric learning problem where a transformation matrix is learned. And the distance is relatively small when extracted features represent the same person and big otherwise. In order to achieve this goal, many effective algorithms have been proposed these years. KISS Metric Learning(KISSME) [9], Cross-view Quadratic Discriminant Analysis (XQDA) [13], Metric Learning with Accelerated Proximal Gradient (MLAPG) [14],Top-push Distance Learning model (TDL) [44] are representative methods. These methods have raised recognition rate to a new level. Besides, Spatially Constrained Similarity function on Polynomial feature map(SCSP) [2] learns similiarity function with Mahalanobis distance and a bilinear similarity metric. Liao et al. [13] proposed a subspace and metric learning method called Cross-view Quadratic Discriminate Analysis (XQDA), which learned a discriminate low dimensional subspace by cross-view quadratic discriminate Analysis and got a QDA metric learned on the derived subspace at the same time. Li Zhang [62] presented to learn a discriminative null space for person re-identification, he used Null Foley-Sammon transform (NFST) to obtain the subspace by collapsing the training data points of each class to a single point. Our feature descriptor can reach high performance based on XQDA and NFST. We will compare our result with other feature descriptors on these two methods.

3 Our approach

For the two key issues of image-based person reidentification, we pay more attention to the first one, That is, the design of robust feature. We have designed a simple, but feasible method to get color and texture feature with high reliability and good computability. What’s more, before extracting feature, We also have proposed to enlarge the image with directional, then we partition it into regions in a special way. In order to better explain the effectiveness of our method, we present the concept of Damaged Region(DR). The flowchart of our method is shown in Fig. 1. The details of each component are as follows.

Fig. 1
figure 1

Flowchart of the proposed method. We first enlarge the images with directional, and partition it into regions. Then we extract features for all images of the specific dataset. Lastly, the image will be split into two sets randomly. The training set is used for distance metric learning, and the testing set uses the measured matrix M which learned by the train set to calculate the match score for each region. The final score will be obtained after comparing the score in our own way

Full size image

3.1 Directional enlargement

As the resolution of the image from existing datasets is relatively low, therefore, we enlarge the image by bilinear interpolation algorithm to increase the useful information. What is different with direct amplification is we adopt directional enlargement. The process of directional enlargement is described in detail in Fig. 2.

Fig. 2
figure 2

The detail of directional enlargement. The top image is enlarged in vertical before partition into regions horizontally. And the same image below is enlarged in horizontal for splitting in vertical

Full size image

The yellow arrow represents the direction of the enlargement, and the red on behalf of the split direction. Before splitting the image in horizontal, we enlarge it in vertical direction with the size of 200 ∗ 80. When partitioning the image into vertical regions, the image is enlarged horizontally and scaled to 140 ∗ 120. Due to the need to take the score of global region into account, we normalize the image to 140 ∗ 80 in the case that the image doesn’t need to be divided.

3.2 Damaged region

In realistic situation, pedestrians usually take something when they walk on the road, such as backpack, bag, book, etc. Due to the different poses and viewpoints, these attachments can’t always be collected by different cameras. This means one of the images will appear some items that do not exist in another. Sometimes, some regions of targets will be covered by these items, which is harmful to identification work. Therefore, in this paper, we name these regions as Damaged Region (DR). For example, Fig. 3 shows some sample images that include Damaged Region from VIPeR dataset. The region marked with yellow wireframe is the Damaged Region.

Fig. 3
figure 3

Examples of Damaged Region. The yellow wireframe in (a) (b) (c) is the DR. (a) Pictures are taken from the front and back. (b) Pictures are taken from the front and side. The DR in (a) and (b) is caused by attachments that captured in second camera. The DR in (c) is caused by the posture and the dress. Images in the same column represent the same person

Full size image

With different viewpoints shown in Fig. 3a shot on the front and back, and Fig. 3b shot on the front and side, we can clearly find that in the second picture there are many new things that do not exist in the first. These new things covered pedestrian’s own useful information, causing interference in matching with the first image. In Fig. 3c, under the influence of the angle and the posture, the dress of the person also emerges variation in local regions, these regions belong to Damaged Region as well.

There is a situation which we have ignored is that the pedestrian attachments appear in both pictures taken by different cameras with different locations, as shown in Fig. 4. Although these items also cover some areas that marked by green wireframe, because of its appearance in two images, these areas turn into the favorable factors for person re-identification, and it does not belong to the Damaged Region. We termed this region in Fig. 4 as Pseudo-damaged Region.

Fig. 4
figure 4

Examples of attachments appear in two images. (a) Attachments cover the region in horizontal. (b) Attachments cover regions in vertical. Images in the same column represent the same person

Full size image

3.3 Partition into regions

After the introduction of the concept of the Damaged Region, we now discuss how to reduce the influence of it. In Fig 1, considering the damaged region is usually rectangular, we divide the picture into rectangular blocks so that the damaged area can be separated. The Rri represented the discrete region, where r is the number of regions and i ∈ [1,r]. The shape and size of the damaged area are also considered, so we have divided the images into horizontal four regions R41, R42, R43, R44, vertical 3 regions R31, R32, R33. And through the experiment, the result is consistent with the analysis.

The Fig. 5 shows the example of how to eliminate the impact of damaged region by dividing the image in horizontal. R41, R42, R43, and R44 are split in horizontal, the corresponding match scores are S41, S42, S43, and S44. We compare the S41 and the S42 to obtain the higher one, and S43 and S44 will be compared at the same time. As shown in Fig. 5, the higher scores are S41 and S43, the homologous regions R41 and R43 are regarded as the valid region, The damaged region R42 usually gets a lower score, and is judged as an invalid one while compared with the others. The final score in horizontal when split into 4 regions S4 is simply represented as:

$$ S_{4} = \left\{ \begin{array}{ll} S_{41}+S_{43} & \text{\(S_{41}>S_{42},S_{43}>S_{44}\)}\\ S_{41}+S_{44} & \text{\(S_{41}>S_{42},S_{43}<S_{44}\)}\\ S_{42}+S_{43} & \text{\(S_{41}<S_{42},S_{43}>S_{44}\)}\\ S_{42}+S_{44} & \text{\(S_{41}<S_{42},S_{43}<S_{44}\)} \end{array} \right. $$
(1)
Fig. 5
figure 5

Illustration of the regional validity judgment when the image is divided into regions in horizontal. The image A is in probe set and the image B is the same person that from gallery set

Full size image

Another important reason for our comparison between the R41 with R42, R43 with R44 is: If there is no damaged region in the image, the first and fourth parts usually contain more background information, which is also a harmful factor in the person re-identification. Through regional validity judgment, the region that contains more background information also be regarded as an invalid region, and not should be added into the final result.

In the vertical direction, as shown in Fig. 3b, the damaged area is usually located on the left or right of the whole image area when it captured from frontal and profile view. In this case, the image is divided into three sub-regions R31, R32, R33. We compare the S31 with S33, then the higher one adds S32 as the final score, yielding vertical score S3:

$$ S_{3} = \left\{ \begin{array}{ll} S_{31}+S_{32} & \text{\(S_{31}>S_{33}\)}\\ S_{33}+S_{32} & \text{\(S_{31}<S_{33}\)} \end{array} \right. $$
(2)

For Pseudo-damaged Region, Obviously, it is not feasible only to split the picture into four blocks in horizontal and three blocks in vertical. Consequently, we take the score of global image S1 into account to compensate for the loss of Pseudo-damaged Regions that have been disabled. Furthermore, People’s coat, whether its style or color, is usually more abundant. Generally, we judge people mostly from their coats, and the identification rate for upper body is more effective than those for lower body. So in order to make full use of the identification effect of the upper body, we will divide the person image into two regions in horizontal and compare the upper body and lower body to get the higher result S2. From what has been discussed above, the overall similarity score S is given by:

$$ S = S_{1}+S_{2}+S_{3}+S_{4} $$
(3)

3.4 Feature extraction

The color is an important feature for describing personal images. In this paper, we describe the object in HSV color space, which is insensitive to changes in light. Besides the color feature, we also apply the Scale Invariant Local Ternary Pattern (SILTP) descriptor for illumination invariant texture description, which is an improved LBP description operator. It is robust to region-wide noise, especially when the detection region is extremely dark or covered by shadows. The descriptor has a strong adaptability.

In the first place, an image is partitioned into many non-overlap horizontal and vertical stripe regions, each region serves as an independent individual. Then we divide each region into a collection of overlapped patches following [13], and extract HSV and SILTP histograms from each patch. We check all patches at the same horizontal location and take the average histogram of each feature among these patches. At last, the average histograms that belong to the same stripe region are connected. The resulting histogram achieves some invariance to viewpoint changes and robustness.

4 Experiments

4.1 Datasets and settings

Datasets

Three widely used datasets are selected for experiments, including VIPeR, PRID and CUHK01.

VIPeR [5]. The VIPeR dataset is one of most challenging datasets for person re-identification task that has been widely used for benchmark evaluation. It contains two views of 632 persons. Each pair for a person is captured by different cameras with different pose, view angle, and illumination conditions. All images are scaled to 128 ∗ 48 pixels. We randomly select 316 persons for training, and the rest persons for testing. We repeat the procedure 10 times to get an average performance.

PRID2011 [7]. The PRID2011 dataset has two camera views as well. Camera A contains 385 persons, whilst camera B captures 749 persons. Only 200 persons appear in both views. We randomly select 100 persons to form the train set, while the remaining 100 persons of camera A to constitute the probe set and the remaining 649 of camera B are used as the gallery, likewise [62]. Experiments are repeated over the 10 times.

CUHK01 [10]. The CUHK01 dataset contains 971 persons from different camera view, and each person has two images in each camera view. Camera A captures the persons in frontal or back views while camera B captures them in the side view. All images were scaled to 160x60 pixels. We randomly divide the dataset under two settings single-shot and multi-shot. For each setting, the training/test split is repeated 10 times and the average of CMC result is reported.

Evaluation metrics. We use Cumulated Matching Characteristics (CMC) curve to evaluate the performance of person re-identification methods for selected datasets in this paper. this is an estimate of finding the correct match in the top n match, which is also known as rank-n. In order to more easily compare with published results, we report the cumulated matching result at selected rank-i (i ∈ 1,5,10,20) in following tables.

4.2 Performance comparison

Our method can flexibly adapt any existing person re-identification methods of feature descriptors and metric learning models. In this section, we achieve our final results based on NFST. For fair comparison, we only compare with the results of using a single type of feature. The bold results in the table represent the best performances at present.

Results on VIPeR

We first evaluate our method against the state-of-the-art on VIPeR. We are only comparing with the classic results in the past five years. Included KISSME [9], Mid-level [63], SCNCD [43], KLFDA [41], Sementic [32], CSL [31], XQDA [13], MLAPG [14], IDLA [1], S-CIR-CNN [35], SSDAL [33], NFST [62], S-SVM [62]. In addition, as paper [1, 33, 35] are the state-of-the-art methods based on deep learning, paper [9, 13, 14, 31, 32, 41, 43, 62, 62, 63] are the state-of-the-art methods based on handcrafted feature and metric learning. The best performance on VIPeR is S-SVM [62]. The comparison results are presented in Table 1. Our rank-1 matching rate 49.05% outperforms the most relevant one LOMO+NFST by 6.77% when a single type of features is used. Our method also improved the Rank-1 matching rate by 6.39% compared to the second best one S-SVM. The results show clearly our approach is superior to all deep learning methods on VIPeR dataset.

Table 1 Experimental evaluations on VIPeR dataset
Full size table

Results on PRID2011

We compare the state-of-the-art results [3, 8, 13, 15, 28, 41, 44, 62] reported on PRID2011 in Table 2. [3] belongs to the deep learning methods, and the rest are based on feature extraction and metric learning. TDL [44] get the best performance on PRID2011. Similar to the performance on VIPeR, the result is improved dramatically when compared with the previous state-of-the-arts. More specifically, we achieved 38.7% of the Rank-1 matching rate, whilst 30.22% for TDL, 29.8% for NFST, and 22.0% for MCP-CNN. Compared with the most related one NFST [62], our results increased by 8.9%, and 16% more than deep learning method MCP-CNN [3]. Comparing with TDL, We get higher results on rank-1 and rank-5, but not on rank-10 and rank-20. This is because TDL used multiple image frames for training and testing, while we just use only one of them for matching.

Table 2 Experimental evaluations on PRID2011 dataset
Full size table

Results on CUHK01

Compared with VIPeR and PRID2011, CUHK01 are much bigger with thousands of training samples. We achieve the final result with single-shot and multi-shot, respectively. Mid-level [63], Semantic [32], IDLA [1], Met-En [28] are state-of-the-art methods with single-shot, while KLFDA [41], KCCA [15], MLAPG [14], NFST [62] are state-of-the-art methods with multi-shot. [13] achieve the result in two respects.

Single-shot

For single-shot, 486 persons are randomly sampled for training and 485 for testing following [13, 26]. Results shown in Table 3 indicates that the proposed approach outperforms other feature descriptors such as Mid-level [63] and LOMO [13]. Our reports the best rank-1 recognition rate of 59.30%, with an improvement more than 10% over LOMO [13] which is regarded as the most relevant one, and an improvement of 11.77% over the deep learning method IDLA. It outperform the state-of-the-art method Met-En [28] with 5.9%, while both of them largely outperform the other existing state of the art methods.

Table 3 Experimental evaluations on CUHK01 dataset with single-shot
Full size table

Multi-shot

For multi-shot, 485 persons are randomly selected for training and 486 for testing following [13, 26, 62]. Multi-shot matching scenario fused scores of multiple images of the same person by the sum rule. We compared available multi-shot results of KLFDA and KCCA from [62]. In Table 4, our approach outperforms NFST by achieving the recognition rate of 70.45%.

Table 4 Experimental evaluations on CUHK01 dataset with multi-shot
Full size table

4.3 Effect of major components

We perform detail analysis of our approach on the VIPeR dataset. The two sets of Probe and Gallery each have 316 person images. The rank-1 indicates that the correct match result is at the top of all results. Therefore, only rank-1 matching rate which is the most important result is selected to compare in the following experiment.

Effect of new feature descriptor

To better illustrate the effectiveness of our new feature descriptor, we compare ours with three classical feature descriptors, including LOMO, GOG, and SCSP. For fair comparison, we only choose one pixel feature vectors of yM𝜃RGB for GOG descriptor. Similarly, we only use one visual cue that composed of color and texture features HSV/SILTP and divided into four stripes for the SCSP. At the same time, we select two methods to achieve the final result, one is NFST which belongs to subspace learning, and the other is XQDA that combine metric learning with subspace learning. The results of three features on XQDA come from the respective papers [2, 13, 26]. The results on NFST are achieved by our own according to the code provided. In Table 5, our feature descriptor evidently improves other feature descriptors both on XQDA and NFST.

Table 5 Comparison with other feature descriptors
Full size table

Effect of partition regions

In order to better illustrate the role of local region, we block the images both in the horizontal and vertical direction. The images are divided equally into regions Rr in both directions, and r= 2, 3, 4, 5, 6. It is necessary to investigate which partition is most effective. Our test is divided into two branches: considering the damaged region(C-DR) and not considering the damaged region (NC-DR). When the damaged area is not considered, we just sum the results of all blocks to get the final result. The Fig. 6 shows the result of Rank-1 matching rate. As shown in (a),(b), the R2 and R4 in horizontal are higher than others, and R3 were significantly higher than those of the other blocks in vertical. What’s more, as the number of r increasing, the recognition rate is gradually reduced both in horizontal and vertical.

Fig. 6
figure 6

Effect of major components. 1.The abscissa represents the number of regions divided, and the ordinate is the Rank-1 matching rate. 2. The blue bar means C-DR, and the yellow bar is NC-DR. 3. (a),(d),(e) in the first row are the results of the horizontal partition, and (b),(d),(f) in second row are the results in vertical. 4. (a), (b) show the results of partition regions in their respective direction. (c),(d) show the results when adding the global region. (e),(f) show the results of horizontal-vertical collaboration

Full size image

Effect of global region

We study the effect of pseudo-damaged regions by observing how the performance changes when the local region combined with global region in each direction. The experiments are also divided into two branches. From the results shown in Fig. 6(b),(c), we can obtain the following observations: (1) R2 and R4 in horizontal and R3 in vertical still get the higher score. This further proves the rationality and effectiveness of our scheme. (2) Compared with (a),(b), we can find that (c),(d) results are significantly improved when taking the score of global region into a count. This is because the pseudo-damaged area is actually a favorable area, when testing the results of each separate partition like (a),(b), we usually treat the pseudo-damaged region as a damaged region, so that the result is damaged too.

Effect of horizontal-vertical collaboration

Combining the partition in vertical and horizontal is a major innovation in the paper. It is necessary to verify the effectiveness of the combination, and how to combine to achieve the best results. Firstly, when testing the effect in horizontal, we keep the images divided into 3 regions in vertical, owing to R3 get the highest result in (a),(b),(c),(d). As before, the R2 regions and the R4 are the best, and much higher than before in Fig. 6e,f. Therefore, when test in vertical, we maintain the picture divided into R2 and R4 in the horizontal direction, both block results are included in the final grade. The results get a much greater improvement over (c),(d). This fully illustrates the effectiveness of the combination of the two directions. In addition, it also explained those R2 and R4 in horizontal combine with R3 in vertical is the best formation, which gets the best Rank-1 matching rate.

Effect of damaged region

In this paper in order to prove the rationality of the damaged area, we carry out the experiments by divided them into two parts of C-DR and NC-DR. Through the previous experimental results in Fig. 6. We can draw the following conclusions:

  1. (1)

    When the effect of each block is verified independently, (a),(b) clearly shows that the result of not considering the damaged region is higher than considered. In this case, compared to the direct added, once considered the damaged area, the role of entire region will be removed and much useful information will be reduced too. After all, only part of whole damaged area is covered, block can not be a complete removal of the specific part.

  2. (2)

    When joining the global region, as shown in (c),(d),(e),(f), the results of C-DR is higher than NC-DR. Due to considering the damaged area and pseudo damaged area at the same time, the inadequacy caused by only considering damaged area will be made up, and improving the recognition rate. This also fully demonstrates the effectiveness of the damaged area and constraints.

  3. (3)

    In all cases, the R2 and R4 in horizontal and R3 in vertical always get the best results when compared with other block results respectively under the same experimental environment. This is consistent with the characteristics of the damaged region analyzed in Sec III.

Effect of directional enlargement

In order to investigate the effect of amplification, we take experiment in four cases:

  1. (1)

    Non-enlargement(NE)

  2. (2)

    Enlarge it in horizontal(EH)

  3. (3)

    Enlarge it in vertical(EV)

  4. (4)

    Directional enlargement(DE)

The results as shown in Table 6, EH performs close to EV, EH and EV evidently improve NE by enlarging the image. DE takes advantages of the three, achieving 49.05% rank-1 matching rate.

Table 6 The results of image is enlarged
Full size table

5 Conclusions

In this paper, we have proposed a novel method by splitting the image into sub-regions in horizontal and vertical after enlarging it. The effectiveness of our method stems from the use of local region in two direction, which reduces the risk of mismatching, increases robustness to occlusion and is more flexible to handle pose variation. Our method also benefits from the global region that is complementary to the damaged regions. Extensive experiments on three benchmarks show that our method achieves the state-of-the-art performance. In the future, we will extend our approach by adopting more local region association strategies and by combining with other types of features and metric learning methods, which is expected to achieve better performance.