Multiple-local feature and attention fused person re-identification method

Abstract

Person re-identification (ReID) is widely used in intelligent security, monitoring, criminal investigation and other fields. Aiming at the problems of local occlusion, scale misalignment and attitude change of pedestrian images in actual scenes, we propose a Multi-local Feature and Attention fused network (MFA) used for person re-identification task. Firstly, Channel Point Affinity Attention module (CPAA) is embedded in the backbone network to enhance the ability of the network for extracting local details. The feature map output from the backbone network is horizontally segmented into four local feature maps, and further four branch networks are concatenated to the feature map of the backbone network. The four local feature maps are used to guide the four branch networks to pay more attention on different areas of pedestrians through Global Local Aligned loss (GLA) function. Finally, the pedestrian feature vector containing multi-local features is obtained. The mAP of the network on Market-1501, DukeMTMC-reID,CUHK03 and MSMT17 datasets were 88.6%, 81.4%, 79.5% and 64.7%, and the Rank-1 was 95.8%, 90.1%, 81.2% and 84.1% respectively. In addition, the model also obtained 73.2% and 68.1% of Rank-1 on partial dataset Patial-REID and Patial-iLIDS, respectively. Recently, The MFA model parameter is 28.3M and the inference efficiency is approximately 32 fps to an image with a resulation of 256 $\times$ 128. Compared with other ReID methods, our proposed methods achieved a competitive performance for ReID task. The code was available at github:git@github.com:ISCLab-Bistu/MFA.git.

Keywords

Person re-identification attention mechanism local feature multi branches network deep learning

1. Introduction

Person Re-identification (ReID) [1] is a technology that can be widely used in video monitoring, intelligent security, and other fields by combining with pedestrian detection, target tracking, and other technologies. It plays an important role in criminal investigation or specific scenarios of person search. The ReID task essentially belongs to the image retrieval problem. In recent years, with the breakthrough of deep learning in the image field, the deep learning-based ReID method has also received widespread attention.

Unlike pedestrian detection tasks, the ReID task needs to pay attention to more granular features of pedestrians. In an open scene, the combined features of pedestrians such as clothing, accessories, and personal belongings are ever-changing. In addition, due to factors such as perspective, posture, occlusion, and lighting, the same pedestrian has multiple differences in different time periods, which makes the ReID task very difficult.

At present, although many pedestrian re-recognition methods have been proposed, it is difficult to solve the problems of background interference and scale variation caused by pedestrian movement. When it is desired to obtain more fine-grained local features, it is easy to introduce background interference, and only learning some features leads to difficulty in improving model performance. In response to these problems, our proposed MFA method can simultaneously address multiple local features while addressing the issues of background interference and pedestrian scale variation. Testing on commonly used datasets Market-1501,Dukemtmc-reid,CUHK03 and MSMT17, as well as occlusion datasets Partial-REID and Partial-iLIDS shows that the network achieves competitive performance. Through visualization analysis using Grad-CAM [27],The proposed method effectively avoids interference from backgrounds and occlusions, allowing for extraction of purer pedestrian features. Furthermore, by incorporating multiple local features of pedestrians, the model’s recognition capability is further improved. When combined with pedestrian detection models, it can efficiently identify target individuals in surveillance videos, providing great convenience for criminal investigation work.

Our work contributions are summarized as follows:

We propose a novel person re-identification method that improves recognition by focusing on multiple local features of pedestrians. Testing results on datasets and occlusion datasets show that our method achieves competitive performance compared to similar methods.

We use multiple modules and the proposed Global Local Aligned (GLA) loss for training. Through GradAM visualization, we find that our method effectively handles complex backgrounds occlusions in pedestrian activity scenes, addressing the problem of background interference commonly encountered in existing methods and providing new insights for future research.

We propose a flexible and effective Channel Point Affinity Attention module (CPAA), which further enhances the model’s performance.

The remaining sections of this paper are organized as follows:

In Section 2, we introduce the related work. In Section 3, we provide detailed explanations of the basic principles and specific implementation of our method. In Section 4, we conduct experimental verification of the method, presenting specific evaluation metrics and comparisons with other methods, as well as visualization results to further validate its effectiveness. Finally, in Section 5, we conclude the paper.

2. Related work

Currently, the ReID task combines research contents of representation learning [2] and metric learning [3], where representation learning enables the network to extract more discriminative features, and metric learning maps features to specific subspaces, increasing the sample class distance and achieving the purpose of distinguishing different pedestrians.

In the ReID task, methods of representation learning mainly include methods based on global features, generative adversarial networks (GANs), pose estimation, masks, image blocks, and attention mechanisms. Among them, networks based on global features such as OSNet [4], BagTricks [5], CDNet [48] and SVDNet [6] nly allow the network to learn partial discriminative features through the dataset and do not consider local features. They perform generally in scenes where pedestrians are occluded. GAN-based methods such as Camstyle [7] and PN-GAN [8] improve the model’s generalization ability by further expanding the training data. Pose estimation-based methods such as CPM [9], GLAD [10], PIE [11], and SpindleNet [12] utilize the key point information of the human body to alleviate the problem of local misalignment. However, it increases the model’s computational complexity and requires human keypoint annotated data during training. Mask segmentation-based methods such as SPReID [13], ISP [51] and MaskReID [14] use image segmentation methods to separate human and background information, generate binary masks, and thus suppress background interference. Image block-based methods such as AlignedReID [15], SCPNet [16], PCB [17], Pyramid [18], BFE [19] and DAAF-BOT [45] force the network to focus on features in different regions during the learning process, enhancing the model’s feature robustness. However, if the pedestrian alignment problem is not solved, it will introduce background interference. Attention mechanism-based methods can selectively improve the feature extraction ability of the network, such as AGW [20], Mancs [21], DuATM [22], HA-CNN [23], Transreid [46], AAFormer [49], DenseFormer [50], etc. The attention mechanism improves the model’s feature extraction ability, but the methods based on attention mechanisms only utilize the global features and discriminative local features of the human body and ignore other minor features, resulting in mediocre performance in occlusion and viewpoint changes. Currently, in supervised ReID tasks, more attention is given to the fusion of global and local information to extract more discriminative pedestrian features.

After extracting discriminative pedestrian features, the recognition performance can be further improved by using metric learning methods that design loss functions to increase inter-class distance and decrease intra-class distance. Commonly used metric learning methods include classification loss [24], triplet loss [25], RA-Loss [47] and quadruplet loss [26]. Classification loss sets the number of IDs in the dataset as the number of classes in the network, feeds the feature vector into a fully connected layer, and calculates cross-entropy loss after Softmax. Triplet loss controls the distance between positive sample pairs and negative sample pairs to be less than a set threshold during training, thus achieving sample clustering within the same class. RA-Loss address the intra-pair variance via exploring the informative relation across pairs. Quadruplet loss adds another set of samples from different classes to triplet loss, and controls the distance between positive sample pairs and other samples from different classes to be less than a set threshold, thus reducing intra-class distance and increasing inter-class distance.

Existing works have difficulty balancing network complexity and fine-grained recognition. Mask-based methods require pixel-level annotations and have high computational costs. Pose estimation-based methods suffer from low efficiency due to complex network structures. Block-based methods fail to address background interference and scale variations. Attention-based methods guide the network to focus on more discriminative features but often neglect other local features. To address these issues, our work balances computational costs with multi-local feature recognition. We use global and local association features to guide the network’s attention to specific regions of the human body, effectively filtering out background interference.

3. Method

Figure 1.

Overall network architecture of MFA. We use ResNet50 as the backbone network and embed CPAA modules in the last three layers. The output feature map is processed through multiple convolutional and GeM modules to obtain multiple global embeddings. Simultaneously, the output feature map is divided into spatially corresponding local embeddings. During training, global and local association learning is achieved through the GLA loss. Each global embedding is learned using Soft-triplet loss, Center loss, and cross-entropy-based ID loss.

The ReID network structure proposed in this article, which integrates local features and attention mechanisms, as shown in Fig. 1. For the ReID task, deep features with rich semantic information are required, but increasing the depth of the network can lead to the problem of vanishing gradients, and the training error also increases accordingly. ResNet proposed by He [28] et al. solved the degradation problem of deep networks, so ResNet50 was used as the backbone network in this article. In order to better utilize the associated information in neighboring regions and extract discriminative pedestrian local semantic features, a Channel Point Affinity Attention (CPAA) module is embedded in the last 3 layers of the network. At the same time, in order to better deal with occlusion, scale changes, and other situations, the output feature map of the backbone network is horizontally divided into multiple regions, and each horizontal region is output as a feature vector through global average pooling (GAP), and corresponding local branches are added. The feature map generated by each local branch is subjected to a GeM [29] layer to produce a feature vector, which is paired with the region feature vector obtained by horizontal segmentation and combined through the proposed Global Local Aligned (GLA) loss function to guide each local branch to learn the features of different regions of the pedestrian. To reduce overfitting, better focus on local detailed features, and increase feature robustness, this article uses cross-entropy loss with LabelSmoothing as the identity loss (ID loss), and adds Soft-triplet loss and Center loss for joint training. It should be noted that batch normalization (BN) is used for the features in ID loss in this article.

3.1 Channel point affiniy attention

Figure 2.

CPAA structure diagram, The part before feature map B employs channel-wise self-attention mechanism, while the part after it utilizes spatial pixel-wise self-attention.

In ReID networks, attention mechanisms can generally promote the network to learn more pedestrian features and suppress irrelevant background information. The proposed Channel Point Affinity Attention (CPAA) in this article can better enable the network to learn associated information in neighboring regions and thus extract more discriminative local semantic features. In Fig. 2, the input feature map of the network ${\textit{{A}}}\in{\mathbb{R}}^{C\times H\times W}$ , where C, $H$ , $W$ denote the channel number, height, and width of the feature map, respectively. Let ${\textit{{A}}}_{i}$ and ${\textit{{A}}}_{j}$ represent the feature maps of the i-th and j-th channels, and the channel correlation matrix ${\textit{{X}}}\in{\mathbb{R}}^{C\times C}$ is shown in the Fig. 2, where the correlation degree $x_{\textit{ij}}$ between channel $x_{i}$ and $x_{j}$ is denoted by :

$\displaystyle x_{\textit{ij}}=\frac{e^{\left({{\textit{{A}}}_{i}\cdot{\textit{% {A}}}_{j}}\right)}}{\mathop{\sum}\nolimits_{j=1}^{C}e^{{\textit{{A}}}_{i}\cdot% {\textit{{A}}}_{j}}},i,j\in\left\{{1,\cdots,C}\right\}$ (1)

Let $\gamma$ be the hyperparameter that controls the correlation degree between two channels, and each channel of the resulting feature map B can be represented as:

$\displaystyle{\textit{{B}}}_{i}=\gamma\mathop{\sum}\limits_{j=1}^{C}\left({x_{% \textit{ij}}{\text{{A}}}_{j}}\right),i\in\left\{{1,\cdots,C}\right\}$ (2)

Subsequently, the feature map B is passed through three $1\times 1$ convolutions to generate feature maps ${\textit{{P}}}_{\textit{{1}}}$ , ${\textit{{P}}}_{\textit{{2}}}$ and ${\textit{{P}}}_{\textit{{3}}}$ , respectively. To reduce the computational cost, the feature maps ${\textit{{P}}}_{\textit{{1}}}$ and ${\textit{{P}}}_{\textit{{2}}}$ are passed through a $1\times 1$ convolution to reduce the number of channels to $C_{1}$ . Similar to the above calculation, the spatial correlation matrix ${\textit{{S}}}\in{\mathbb{R}}^{N\times N}$ is obtained, where $N=H\times W$ . Then, the matrix after transformation of S is multiplied with ${\textit{{P}}}_{\textit{{3}}}$ to obtain the feature map F with high semantic correlation on the spatial level. Finally, to alleviate the problem of gradient vanishing, the feature map G is obtained through a residual structure.

3.2 Local branch structure

In this paper, the output feature map of the backbone network is horizontally divided into $R$ regions, and the backbone feature map is output to $R$ parallel local branches through $R$ groups of 1 $\times$ 1 convolution layers. Each local branch is jointly trained with the corresponding horizontally divided feature vector, and the GLA loss function is used to guide each branch network to focus on different local features of the image, and finally fuse multiple local feature vectors. The local branch in Fig. 1 uses the GeM module when generating feature vectors. Compared with global max pooling and global average pooling, the GeM module has the ability of adaptive learning, which can better retain discriminative fine texture features. The feature vector output by GeM can be represented by the formula:

$\displaystyle{\textit{{f}}}=\left[{f_{1},f_{2},\cdots,f_{K}}\right]^{T}$ (3)

where $K$ represents the number of features, and $f_{k}$ represents the $k$ -th generated feature vector, which can be represented as:

$\displaystyle f_{k}=\left({\frac{1}{\left|{\chi_{k}}\right|}\mathop{\sum}% \nolimits_{x_{i}\in\chi_{k}}x_{i}^{p_{k}}}\right)^{\frac{1}{p_{k}}}$ (4)

In the formula, $\chi_{k}$ represents the input feature map parameter, $p_{k}$ is a learnable parameter, and the partial derivative of $f_{k}$ with respect to $x_{i}$ and $p_{k}$ is represented as:

$\displaystyle\frac{\partial f_{k}}{\partial x_{i}}=\frac{1}{\left|{\chi_{k}}% \right|}f_{k}^{1-p_{k}}x_{i}^{p_{k}-1}$ (5) $\displaystyle\frac{\partial f_{k}}{\partial p_{k}}=\frac{f_{k}}{p_{k}^{2}}% \left({\textit{ln}\frac{\left|{\chi_{k}}\right|}{\mathop{\sum}\nolimits_{x\in% \chi_{k}}x^{p_{k}}}+p_{k}\frac{\mathop{\sum}\nolimits_{x\in\chi_{k}}x^{p_{k}}% \textit{lnx}}{\mathop{\sum}\nolimits_{x\in\chi_{k}}x^{p_{k}}}}\right)$ (6)

Different from global average pooling that only performs pooling operation on the feature map, GeM continuously updates the final convergence to the best value during the backpropagation process when generating the feature vector, and when $p_{k}$ approaches $\infty$ , GeM is equivalent to global max pooling, while when $p_{k}$ is 1, GeM is equivalent to global average pooling. GeM can perform the best pooling operation through adaptive learning.

After generating the feature vector in the local branch, batch normalization (BN) [5] module is applied to normalize the features. Batch normalization makes the features of different input samples follow a normal distribution, which helps the ID loss to converge faster and also facilitates the convergence of other auxiliary losses.

3.3 MFA Network loss function

The overall loss function of the MFA network can be represented as follows:

$\displaystyle{\cal L}_{\text{total}}={\cal L}_{\text{ID}}+\lambda{\cal L}_{% \text{Soft-Triplet}}+\mu{\cal L}_{\text{Center}}+\gamma{\cal L}_{\text{GLA}}$ (7)

In the formula, $\lambda$ , $\mu$ , $\gamma$ are weight factors which can be learned by the network. The BN module in Fig. 1 is only responsible for optimizing the feature for the classification loss part. Therefore, during the module training process, except for ${\cal L}_{\text{ID}}$ using the feature vector after the BN module, other loss functions use the feature vector before BN. During testing, the concatenated feature vector from the local branches is used.

Firstly, the model uses cross-entropy loss with added label smoothing as the basic classification loss function, which can be represented by the formula:

$\displaystyle{\cal L}_{\text{ID}}=\mathop{\sum}\limits_{i=1}^{N}-q_{i}\ln\left% ({p_{i}}\right)$ (8) $\displaystyle q_{i}=\left\{{{\begin{array}[]{ll}1-\frac{N-1}{N}\varepsilon,&i=% y\\ \hbox{${\varepsilon}$}\!\mathord{\left/{\vphantom{{\varepsilon}N}}\right.\kern% -1.2pt}\!\lower 3.01pt\hbox{$N$},&i\neq y\\ \end{array}}}\right.$ (9)

where $p_{i}$ represents the probability value of predicting class i, $y$ represents the true class, $N$ represents the total number of ID categories, and $\varepsilon$ is a small value (0.1 in this paper). Label Smoothing reduces the confidence of the network when predicting correctly, while increasing the confidence of the network when predicting incorrectly, thereby reducing overfitting of the network on the training set.

In addition, to improve the network’s learning of finer-grained features, the Soft-Triplet loss [30] is added, where $d_{a,p}$ and $d_{a,n}$ represent the distances between positive sample pairs and negative sample pairs, respectively. The specific expression of Soft-Triplet loss is:

$\displaystyle{\cal L}_{\text{Soft-Triplet}}=\text{ln}\left({1+e^{d_{a,p}-d_{a,% n}}}\right)$ (10)

The problem with triplet-based losses, including the Soft-Triplet loss, is that the function only focuses on the difference between $d_{a,p}$ and $d_{a,n}$ , but not on the magnitude of $d_{a,p}$ . For example, when $d_{a,p}$ is 0.3, $d_{a,n}$ is 0.5, the loss is equal to $d_{a,p}$ is 1.3, $d_{a,n}$ is 1.5. To address this issue, the Center Loss is introduced [31], which is used to measure the distance between the current class and its intra-class center. The expression is:

$\displaystyle{\cal L}_{\text{Center}}=\frac{1}{2}\mathop{\sum}\limits_{j=1}^{B% }\left\|{{\textit{{f}}}_{t_{j}}-{\textit{{c}}}_{y_{j}}}\right\|_{2}^{2}$ (11)

In the equation, $B$ represents the mini-batch size, ${\textit{{c}}}_{y_{j}}$ represents the feature center of class $y_{j}$ , $y_{j}$ represents the label of the $j$ -th image in the mini-batch, and ${\textit{{f}}}_{t_{j}}$ represents the feature vector of the $j$ -th image. Minimizing the Center Loss can reduce the intra-class variance. Finally, for the various local branches in the network, their own features are generated by the $1\times 1$ convolution layer on the output of the backbone network. The network cannot automatically focus on different local features of pedestrians. Therefore, the GLA loss function is constructed, and the $R$ horizontally segmented region feature maps are used to guide the local branches to learn different local features. For the local feature vectors and their corresponding region feature vectors, the closer the distance between the two, the closer the features that are attended to. Therefore, the loss function is constructed as follows:

$\displaystyle{\cal L}_{\text{GLA}}=\mathop{\sum}\limits_{i=1}^{R}F\left({{% \textit{{f}}}_{\textit{gi}},{\textit{{f}}}_{\textit{li}}}\right)$ (12) $\displaystyle F\left({{\textit{{f}}}_{\textit{gi}},{\textit{{f}}}_{\textit{li}% }}\right)=\left\{{{\begin{array}[]{ll}0.5\ast\left|{{\textit{{f}}}_{\textit{gi% }}-{\textit{{f}}}_{\textit{li}}}\right|^{2},&\quad\left|{{\textit{{f}}}_{% \textit{gi}}-{\textit{{f}}}_{\textit{li}}}\right|<1\\ \left|{{\textit{{f}}}_{\textit{gi}}-{\textit{{f}}}_{\textit{li}}}\right|-0.5,&% \quad\left|{{\textit{{f}}}_{\textit{gi}}-{\textit{{f}}}_{\textit{li}}}\right|% \geqslant 1\\ \end{array}}}\right.$ (13)

In the formula, presents the number of horizontally divided regions, ${\textit{{f}}}_{\textit{gi}}$ represents the semantic features included in the $i$ -th local branch, ${\textit{{f}}}_{\textit{li}}$ represents the $i$ -th split region feature. Therefore, ${\textit{{f}}}_{\textit{gi}}$ is supervised by ${\textit{{f}}}_{\textit{li}}$ to make the f-th local branch tend to focus on the local information of the $i$ -th region. At the same time, since the output features of the local branch contain global semantic information, when there are difficulties such as non-aligned pedestrian scales or pedestrian occlusion, i.e., the $i$ -th region does not contain the required pedestrian features to be focused on, but only contains background interference information, the semantic features cannot match the corresponding local features, so the local features of this region will be ignored, thus achieving that different local branches only focus on pedestrian local features and eliminate the influence of background interference.

To provide a clearer representation of the implementation details, the pseudocode is as follows:

Training Prodecure of MFA
Input:
Training data $D_{\textit{train}}$ ; Evaluation data $D_{\textit{eval}}$ ; MaxIters; MaxEpoches.
Output:
Feature extractor F( $\cdot$ ); local branches
1	Initialization learning rate lr;
2	for epoch $=$ 1 to MaxEpoches do
3	for iter $=$ 1 to MaxIters do
4	Space wise F and pooling into four local embeddings: L1, L2, L3, L4;
5	Channel wise F and pooling into four global embeddings: G1, G2, G3, G4;
6	Update F( $\cdot$ ) by ${\cal L}_{\textit{Soft-triplet}},{\cal L}_{\textit{Center}}$ and ${\cal L}_{\textit{ID}}$ .
7	Update local branches by ${\cal L}_{\textit{GLA}}$ with L1&G1, L2&G2, L3&G3, L4&G4.
8	end
9	evaluation $D_{\textit{eval}}$ to save the best model.
10	end

4. Experiments

4.1 Experimental dataset introduction

Table 1
Parameters of the datasetsThe datasets used in our experiments include market1501, DukeMTMC-reID CUHK03, and MSMT17, which consist of both training and testing sets. Partial-REID and Partial-iLIDS datasets are used for testing occluded scenarios only

DataSet	#identities	#imgs	Train #imgs	Train #identities	Test #identities	Gallery #imgs	Query #imgs	#Camera
Market1501	1501	32668	12936	751	750	16364	3368	6
DukeMTMC-reID	1404	36411	16522	702	702	17661	2228	8
CUHK03	1467	13164	7368	767	700	5328	1400	10
MSMT17	4101	126441	32621	1041	3060	82161	11659	15
Partial-REID	60	600	–	–	60	300	300	–
Partial-iLIDS	119	476	–	–	119	238	238	–

Figure 3.

Examples of cropped pedestrian images on 6 different datasets where (e) and (f) are occlusion datasets.

To demonstrate the effectiveness of our proposed method, we conducted experiments on four widely-used datasets, Market-1501, DukeMTMC-reID, CUHK03 and MSMT17, as well as two occlusion datasets, Partial-REID and Partial-iLIDS, to evaluate the model’s generalization ability under occlusion scenarios. Figure 3 shows some examples of each dataset. Market-1501, DukeMTMC-reID, CUHK03 and MSMT17 were all collected in university campus environment and contain a certain amount of occlusion; Partial-REID was captured outdoors, where pedestrians are occluded by trees, buildings, and other objects; Partial-iLIDS was captured by airport cameras, where pedestrians are occluded by luggage, signs, and other pedestrians. Table 1 shows the basic parameters of each dataset, where #identities represents the number of identities, #imgs represents the number of images, and #camera represents the number of cameras. Partial-REID and Partial-iLIDS are relatively small in scale and do not include a training set, only being used for testing purposes.

4.2 Experimental configuration

The experiment was conducted on Ubuntu20.04 operating system with hardware specifications including core-i9 11900 CPU and NVIDIA GeForce 3090 graphics card with 24 GB of memory. The model was built using Pytorch framework and the ResNet50 network pre-trained on the ImageNet dataset was used as the backbone network. Since the pedestrian images in the dataset are generally small in size and the feature map resolution is low after downsampling by 32 times, the stride of the last layer of the backbone network was set to 1. During training, the Adam optimizer was used with a batch size of 64 (16 $\times$ 4) and RandomIdentitySampler was used for sampling. The initial learning rate was set to 3.5 $\times$ 10^{- 4} and weight decay was set to 5 $\times$ 10^{- 4}. A total of 120 epochs were trained with a warm-up mechanism for the first 40 epochs, gradually increasing the learning rate to 3.5 $\times$ 10^{- 4}. The learning rate was then decreased to 3.5 $\times$ 10^{- 5} and 3.5 $\times$ 10^{- 6} at 40 and 70 epochs, respectively. The image resolution was scaled to 256 $\times$ 128, and data augmentation techniques such as horizontal flipping and random cropping were used with probabilities of 0.5 each. Moreover, when the network input resolution is 256 $\times$ 128 and is downsampled by 16 times, the output is a feature map of size 16 $\times$ 8. To ensure that the semantic information of the regions is not lost,also enable the localbranches take better effect,finally, the feature map was divided into 4 parts, covering the four regions of the body (above the chest, from chest to waist, from waist to knee, below the knee), which can effectively capture each local feature of pedestrians on most ReID datasets. During testing, the pedestrian feature vectors were L2 normalized, and the model performance was evaluated using CMC (Cumulative Matching Characteristics) and mAP (mean Average Precision) metrics. For each query, one image was selected at a time and the similarity with all images in the gallery was calculated. The CMC metric records the probability of successful matching for the top $k$ images, and the probability of the first successful match is recorded as Rank-1. Assuming that the model ranks candidates based on similarity from highest to lowest, the Rank-k metric is calculated using formula:

$\displaystyle\textit{Rank-}k=\frac{1}{N}\sum\limits_{i=1}^{N}{p_{i}}$ (14)

$N$ represents the number of queries, where each query examines the top $k$ candidates. If the ground truth is included among the candidates, $p_{i}$ is assigned a value of 1; otherwise, it is assigned 0.

Considering a model’s prediction, True Positive (TP) refers to predicted positive samples, False Positive (FP) indicates incorrectly predicted positive samples, True Negative (TN) represents correctly predicted negative samples, and False Negative (FN) denotes incorrectly predicted negative samples. Precision is denoted as:

$\displaystyle\textit{Precision}=\frac{\textit{TP}}{\textit{TP}+\textit{FP}}% \times 100\%$ (15)

Consider an ID labeled as $i$ , with $N_{i}$ representing the query images associated with it. The number of queries is $N_{i}$ , and the average precision (AP) is the mean precision of each query, denoted as:

$\displaystyle\textit{AP}=\frac{\sum{\textit{Precision}_{i}}}{N_{i}}$ (16)

mAP represents the mean average precision across all IDs, with $M$ denoting the total number of IDs. The mAP formula is defined as :

$\displaystyle\textit{mAP}=\frac{\sum\limits_{i=0}^{M}{\textit{AP}_{i}}}{M}$ (17)

4.3 Ablation experiment

Table 2
Effects of CPAA module position on model performance the CPAA module in the 2nd and 3rd layers of the backbone network leads to significant improvement in metrics.adding the CPAA module in the last layer still has a positive effect

Method	Baseline	Layer2	Layer3	Layer4	Market-1501		DukeMTMC-reID
					Rank-1	mAP	Rank-1	mAP
1	$\surd$				93.6	85.1	85.4	76.1
2	$\surd$	$\surd$			94.2	85.4	85.9	76.5
3	$\surd$	$\surd$	$\surd$		95.0	86.6	86.8	78.1
4	$\surd$	$\surd$	$\surd$	$\surd$	95.2	86.8	87.1	78.2

To investigate the optimal location of the CPAA module in the network, four groups of control experiments shown in Table 2 were designed. Baseline only uses the ResNet50 backbone network to extract features without multiple local branches and the CPAA attention structure, and the loss function does not contain GLA. Layer2, Layer3, and Layer4 represent adding the CPAA module at the corresponding positions. To ensure the fairness of the comparison, the basic experimental parameters were set according to Section 4.2, and each method was trained and tested on the Market-1501 and DukeMTMC-reID datasets separately.

Table 2 shows the results of four control experiments designed to study the optimal location of the CPAA module in the network. Baseline only uses the ResNet50 backbone network to extract features, without including multiple local branches or the CPAA attention structure, and the loss function does not include GLA. Layer2, Layer3, and Layer4 represent adding the CPAA module at the corresponding positions. To ensure fair comparison, the basic experimental parameters are set according to Section 3.2, and each method is trained and tested on the Market-1501 and DukeMTMC-reID datasets. The data in Table 2 shows that using only the Baseline method 1, the Rank-1 indicators in Market-1501 and DukeMTMC-reID are 93.6% and 85.1%, respectively. Methods 2, 3, and 4 all show some degree of improvement in Rank-1 indicators, because the CPAA module aggregates channel features and related information from neighboring regions, resulting in feature maps with better representation capabilities. Method 3 shows a higher degree of improvement compared to Method 2, as the CPAA module in Method 2 is located in a shallower position in the network, with more interference from other low-level features. In Method 3, the feature map outputted after Layer3 already contains rich semantic features, and at this point, adding the module can better aggregate deep semantic features and improve network performance. Method 4 shows some improvement compared to Method 3, but the magnitude of improvement is limited, as important semantic information has already been attended to through multiple feature aggregations in the early stages, and the outputted features tend to be stable.

To further verify the effectiveness of each module, six control experiments are designed, as shown in Table 3, where Localbranches represents multiple local branches.

Table 3

Effects of different module combinations on model performanceon the Market-1501 and DukeMTMC-reID datasets, CPAA directly improves detection metrics, and further improvements are observed with the influence of GLA and local branches. Despite scale variations and background interference, each local branch focuses on the corresponding pedestrian region and ignores the background area

Method	Market-1501		DukeMTMC-reID
	Rank-1	mAP	Rank-1	mAP
Baseline	93.6	85.1	85.4	76.1
Baseline $+$ CPAA	95.2	86.8	87.1	78.2
Baseline $+$ Localbranches	93.5	85.2	85.3	76.0
Baseline $+$ Localbranches $+$ CPAA	95.3	86.7	87.3	78.5
Baseline $+$ Localbranches $+$ GLA	95.0	87.3	88.6	80.2
MFA (ours)	95.8	88.6	90.1	81.4

A comparative analysis of the data in Table 3 reveals the following: (1) CPAA, when added to Baseline, increases the Rank-1 performance by 1.6% and 1.7% on the Market-1501 and DukeMTMC-reID datasets, respectively. Similar improvements are observed in other combinations that include CPAA, indicating that CPAA can effectively aggregate semantic features from neighboring regions in the feature map and enhance the representation ability of the backbone network. (2) Compared to Baseline, adding only Localbranches (without CPAA or GLA) to Baseline has no significant impact on the Rank-1 performance on both datasets, and this observation holds for other combinations that only include Localbranches, indicating that Localbranches alone cannot achieve the goal of diversifying local features. (3) Adding GLA to Localbranches (Baseline $+$ Localbranches $+$ GLA) increases the Rank-1 performance by 1.5% and 3.3% on the Market-1501 and DukeMTMC-reID datasets, respectively, compared to Baseline $+$ Localbranches, suggesting that Localbranches can extract multiple local features and obtain feature vectors that contain more detailed information only when used with GLA in joint training. Based on the above analysis, it can be concluded that CPAA and Localbranches combined with GLA have the potential to improve the model’s performance. Finally, the MFA model is built by stacking the above modules, and achieves Rank-1 performance of 95.8% and 90.1% on the Market-1501 and DukeMTMC-reID datasets, respectively, representing a 2.2% and 4.7% improvement over Baseline.

4.4 Network visualization analysis

Figure 4.

Grad-CAM results of MFA, with the influence of the GLA and local branches, even if the pedestrian scale changes and there is background interference, each local branch focuses on the corresponding pedestrian area and ignores the background area.

To visualize the regions of interest in the image that the network attends to, Grad-CAM [27] was used for visual analysis. Four representative scene images were randomly selected from the Market-1501 test dataset, as shown in Fig. 4. The leftmost image in each group is the original image, and the middle four images show the visualization of the network’s attention regions overlaid with four different local branches. The image on the far right shows the visualization of the fused attention region after combining the four local branches.

In Fig. 4, (a) shows a normal cropped image, as described in Section 4.2. The different branches of the network sequentially attend to four regions of the pedestrian’s body, from above the chest to the chest and waist, from the waist to the knees, and from below the knees. The fused attention regions basically cover the entire outline of the pedestrian. At the same time, the fused features tend to focus on more discriminative regions, such as special texture patterns on the upper body, and the transition area between the lower body pants and legs. (b) shows a situation where the scale is changed, and the pedestrian only occupies a local area of the image. In this case, the local branch responsible for the upper part of the image does not attend to the background area above the image because under the influence of global semantic features, there is no semantic information that meets the requirements in the upper region of the image. This branch of the network defaults to outputting the most discriminative global feature, i.e., the upper body region of the pedestrian. The other branches continue to focus on different local regions. (c) shows a situation where the pedestrian is partially occluded. Because the lower part of the pedestrian’s body is covered, the branch network responsible for the area below the knees defaults to outputting the most prominent global feature, and the other branches only focus on the upper body region. (d) shows a difficult situation where there is both background interference and scale change. Many other pedestrians and objects appear in the background, but the attention regions of each local branch still meet the requirements.

In all four situations, the network’s final attention regions cover the entire pedestrian and ignore interference such as occlusion and background, thus achieving rich and pure pedestrian features in the network output.

Figure 5.

Grad-CAM results of MFA on Patial-REID dataset images, The model trained on the market1501 dataset still pays attention to the pedestrian region and ignores the occlusion area when tested on the Patial-REID dataset.

To observe the generalization ability of the network on other occlusion datasets, we selected four representative images from the Patial-REID dataset, as shown in Fig. 5, including half-body, side view, heavily occluded, and partially occluded situations. After visualizing with Grad-CAM, it was found that the model trained on the Market-1501 dataset still has good generalization ability on Patial-REID. The network’s attention areas covered the entire person and focused on multiple discriminative local features while ignoring the interference of the background and occlusions.

4.5 Comparision with other methods

The proposed method will be compared with current state-of-the-art (SOTA) methods, including block-based methods such as DAAF-BOT [45], MGN [32], GCP [33], MSCN [34]; global-based methods such as OSNet [4] and BagTricks [5],CDNet [48], AEFLN [35]; mask-based segmentation method such as ISP [51], SPReID [13], GASM [36]; attention-based methods such as AAFormer [49], DenseFormer [50], AGW [20], ABD-Net [37], and SONA [38]. To ensure fairness in comparison, re-ranking [39], multi-query [40], and other methods will not be used in this study.

Table 4
Performance comparison of different methods on 4 datasetsThe evaluation metrics are shown in bold, while underline indicates the second-best result. ‘–’ denotes that the result is unavailable. Our model achieves competitive performance across the evaluated datasets

Method	Market-1501		DukeMTMC-ReID		CUHK03		MSMT17
	Rank-1	mAP	Rank-1	mAP	Rank-1	mAP	Rank-1	mAP
OSNet [4]	94.8	84.9	88.6	73.5	72.3	67.8	78.7	52.9
BagTricks [5]	94.5	85.9	86.4	76.4	–	–	–	–
DAAF-BOT [45]	95.1	87.9	87.9	77.9	69.0	67.6	–	–
TransReID [46]	94.9	88.1	90.2	81.3	–	–	83.5	65.5
RA-Loss [47]	95.7	86.6	90.6	79.8	76.6	70.6	–	–
CDNet [48]	95.1	86.0	88.6	76.8	–	–	78.9	54.7
AAFormer [49]	95.4	87.7	90.1	80.0	79.9	77.8	83.6	63.2
Denseformer [50]	94.4	87.3	89.0	79.8	–	–	82.2	62.6
ISP [51]	95.3	88.6	89.6	80.0	76.5	74.1	–	–
MGN [32]	95.7	86.9	88.7	78.4	68.0	67.4	–	–
GCP [33]	95.2	88.9	89.7	78.6	77.9	75.6	–	–
ABD-Net [37]	95.6	88.3	89.0	78.6	–	–	–	–
SONA [38]	95.7	88.7	89.3	78.1	81.8	79.2	–	–
MSCN [34]	94.2	86.5	88.1	77.9	77.6	75.2	–	–
AEFLN [35]	93.3	81.3	86.5	72.9	66.5	61.7	–	–
GASM [36]	95.3	84.7	88.3	74.4	–	–	79.5	52.5
AGW [20]	95.1	87.8	89.0	79.6	63.6	62.0	68.3	49.3
MFA(Ours)	95.8	88.6	90.1	81.4	81.2	79.5	84.1	64.7

Table 4 presents the performance of these networks on the Market-1501,DukeMTMC-reID,CUHK03 and MSMT17 datasets. OSNet, AGW, and MGN, which have better performance among the methods described above with open source,were selected for analysis. OSNet incorporated a fusion mechanism of convolution layers with different receptive fields in the module, which can extract the most discriminative features, but ignores other secondary features and has limited network generalization ability. AGW further enhances the ability of the network backbone to extract features by adding an attention mechanism module on the basis of BagTricks, but it does not use local feature information. MGN uses global and block-based local features, and the network has the ability to extract fine-grained features, but it does not address the problem of scale transformation and occlusion, which may introduce more interference information when the network faces such situations. In this paper, by using CPAA and multi-branch structures, the network can extract diverse fine-grained features, and by using GLA, the network considers global semantic information while focusing on local features to deal with scale transformation and occlusion scenarios. The indicators in Table 4 show that the proposed MFA network has the best performance.

Figure 6.

The top-10 detection for each model under the same query are shown, with red indicating incorrect ID matches and green indicating correct ID matches.

To visually reflect the effectiveness of the proposed method, the retrieval results of AGW, OSNet, MGN, and the proposed method were compared. Figure 6 shows the Rank-10 retrieval results of each method and the proposed method, with dotted boxes indicating incorrect matches. Although AGW and OSNet had correct Rank-1 matches, they could not pay attention to more detail information due to the absence of local features, leading to more incorrect matches. MGN failed to solve the problem of scale variation and could result in misidentification in such cases. MFA, due to its ability to extract more diverse local fine-grained features and address the issue of scale variation, achieved better Rank-10 matching.

Figure 7.

For visualization purposes, we compare our results by Grad-CAM. As shown in (a) and (b), from left to right: the original image, RA-Loss [47] results, ISP [51] results, AGW [20] results, and our results. Our model accurately focuses on more discriminative pedestrian features while filtering out background and occlusion interference.

To visually demonstrate the effectiveness of our proposed method in capturing detailed information in images, we compared it with other methods RA-Loss, ISP, and AGW using Grad-CAM visualization. Figure 7 presents the specific visualization results. Our method demonstrates more effective attention to multiple local features in the pedestrian region while filtering out background and occlusion interference.

To compare model complexity and execution efficiency, we selected models that performed well on multiple datasets for comparative analysis. The models include AGW, GASM, ISP, DAAF-BOT, and MGN. Table 5 the parameters and execution efficiency (fps) for each model, with a test image resolution of 256 $\times$ 128. To enhance the feature extraction capability of the network, ResNet50 is used as the backbone network, resulting in a higher number of parameters. Compared to other models, our model has advantages in terms of parameter size and inference efficiency while maintaining accuracy.

Finally, to verify the model’s generalization ability under various occlusion scenarios, tests were conducted on the Partial-REID and Partial-iLIDS datasets using the single-shot method. The occluded image was used as the query and the complete pedestrian image was used as the gallery. Table 6 shows the indicators for each method. In Partial-REID, MFA achieved a Rank-1 indicator of 73.2%, which is 3.5% higher than AGW, ranked higher. In the Partial-iLIDS dataset, the Rank-1 indicator of MFA increased by 0.8% compared to the better-performing PGFA, reaching 68.1%.

Table 5

Comparison of different models in terms of parameter size and inference efficiency

Method	Params(M)	FPS(/s)
AGW [20]	26.7	35
GASM [36]	43.5	21
ISP [51]	28.5	30
DAAF-BOT [45]	33.9	27
MGN [32]	36.8	24
MFA(Ours)	28.3	32

Table 6

Performance comparison of different methods on 2 partial datasetswhile underline represents the second-best. Our model still performs the best on the occlusion dataset containing only test images

Method	Partial-REID		Partial-iLIDS
	Rank-1	Rank-3	Rank-1	Rank-3
DSR [39]	50.7	70.0	58.8	67.2
SFR [38]	56.9	78.5	63.9	74.8
VPM [40]	67.7	81.9	65.5	74.8
PGFA [41]	68.1	79.4	67.3	76.7
AGW [20]	69.7	80.0	64.7	79.8
DAAF-BOT [45]	60.7	68.7	53.8	72.3
MFA(ours)	73.2	83.5	68.1	82.6

5. Conclusion

We propose a pedestrian re-identification network that integrates local features with attention mechanisms. The CPAA attention module is embedded in the ResNet50 backbone network, effectively aggregating semantic features from different channels and neighboring regions. On the other hand, through the joint action of local branches and GLA loss, the model can focus on multiple local features of pedestrians. Visual analysis shows that the fused pedestrian features are rich and pure. Compared with similar networks, the proposed method achieves competitive performance, demonstrating its effectiveness.However, our work still has room for improvement. Currently, pedestrian re-identification models have high accuracy but large parameter sizes, making it challenging to deploy them on edge devices.In future work, we will investigate how to streamline the model while maintaining current performance, enabling lightweight and edge deployment of the network.

Footnotes

Acknowledgments

This research was funded by the National Natural Science Foundation of China (NSFC) (grant no. U21A6003), the Program of Promoting the Development of University – Diligence Talents (grant no. 5112111145).

References

Zhao

Xiao

et al., Deepreid: Deep filter pairing neural network for person re-identification, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2014), 152–159.

Bengio

Courville

and Vincent

, Representation learning: A review and new perspectives, IEEE Transactions on Pattern Analysis and Machine Intelligence 35(8) (2013), 1798–1828.

Kulis

, Metric learning: A survey, Foundations and Trends® in Machine Learning 5(4) (2013), 287–364.

Zhou

Yang

Cavallaro

et al., Omni-scale feature learning for person re-identification, Proceedings of the IEEE, CVF International Conference on Computer Vision (2019), 3702–3712.

Luo

Liao

et al., Bag of tricks and a strong baseline for deep person re-identification, Proceedings of the IEEE, CVF Conference on Computer Vision and Pattern Recognition Workshops (2019).

Sun

Zheng

Deng

et al., Svdnet for pedestrian retrieval, Proceedings of the IEEE International Conference on Computer Vision (2017), 3800–3808.

Zhong

Zheng

et al., Camstyle: A novel data augmentation method for person re-identification, IEEE Transactions on Image Processing 28(3) (2018), 1176–1190.

Qian

Xiang

et al., Pose-normalized image generation for person re-identification, Proceedings of the European Conference on Computer Vision (ECCV) (2018), 650–667.

Zheng

Huang

Y.J.

H.C.

et al., Poseinvariant embedding for deep person re-identification, IEEE Transactions on Image Processing 28(9) (2019), 4500–4509.

10.

Wei

Zhang

Yao

et al., Glad: Global-local-alignment descriptor for pedestrian retrieval, Proceedings of the 25th ACM international conference on Multimedia (2017), 420–428.

11.

Zheng

Huang

et al., Pose-invariant embedding for deep person re-identification, IEEE Transactions on Image Processing 28(9) (2019), 4500–4509.

12.

Zhao

Tian

Sun

et al., Spindle net: Person re-identification with human body region guided feature decomposition and fusion, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017), 1077–1085.

13.

Kalayeh

M.M.

Basaran

Gökmen

et al., Human semantic parsing for person re-identification, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018), 1062–1071.

14.

Huo

Wang

et al., Maskreid: A mask based deep ranking neural network for person re-identification, arXiv preprint arXiv:180403864, (2018).

15.

Zhang

Luo

Fan

et al., Alignedreid: Surpassing human-level performance in person re-identification, arXiv preprint arXiv:171108184, (2017).

16.

Fan

Luo

Zhang

et al., Scpnet: Spatial-channel parallelism network for joint holistic and partial person re-identification, Asian Conference on Computer Vision, Springer, Cham, (2018), 19–34.

17.

Sun

Zheng

Yang

Tian

and Wang

, Beyond part models: Person retrieval with refined part pooling, in ECCV, (2018), 480–496.

18.

Zheng

Deng

Sun

Jiang

Guo

Huang

and Ji

, Pyramidal person re-identification via multi-loss dynamic training, in CVPR, (2019), 8514–8522.

19.

Dai

Z.Z.

Chen

M.Q.

Zhu

S.Y.

and Tan

, Batch feature erasing for person re-identification and beyond, arXiv preprint arXiv:181107130, (2018).

20.

Shen

Lin

et al., Deep learning for person re-identification: A survey and outlook, IEEE Transactions on Pattern Analysis and Machine Intelligence 44(6) (2021), 2872–2893.

21.

Wang

Zhang

Huang

et al., Mancs: A multi-task attentional network with curriculum sampling for person re-identification, Proceedings of the European Conference on Computer Vision (ECCV) (2018), 365–381.

22.

Zhang

C.G.

et al., Dual attention matching network for context-aware feature sequence based person re-identification, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018), 5363–5372.

23.

Zhu

and Gong

, Harmonious attention network for person re-identification, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018), 2285–2294.

24.

Zhang

and Sabuncu

, Generalized cross entropy loss for training deep neural networks with noisy labels, Advances in Neural Information Processing Systems (2018), 31.

25.

Hermans

Beyer

and Leibe

, In defense of the triplet loss for person re-identification, arXiv preprint arXiv:170307737, (2017).

26.

Chen

W.H.

Chen

X.T.

Zhang

J.G.

et al., Beyond Triplet Loss:A Deep Quadruplet Network for Person Re-identification, Proc of the IEEE Conference on Computer Vision and Pattern Recognition, Washington, USA: IEEE, (2017), 1320–1329.

27.

Radenovic

Tolias

and Chum

, Fine-tuning cnn image ’retrieval with no human annotation, IEEE TPAMI 41(7) (2018), 1655–1668.

28.

Zhang

Ren

et al., Deep residual learning for image recognition, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2016), 770–778.

29.

Zheng

Deng

Sun

et al., Pyramidal person re-identification via multi-loss dynamic training, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA: CVPR, (2019), 8514–8522.

30.

Hermans

Beyer

and Leibe

, In defense of the triplet loss for person re-identification, arXiv preprint arXiv:170307737, (2017).

31.

Wen

Zhang

et al., A discriminative feature learning approach for deep face recognition, European Conference on Computer Vision. Springer, Cham, (2016), 499–515.

32.

Wang

Yuan

Chen

and Zhou

, Learning discriminative features with multiple granularities for person reidentification, in ACM MM, (2018), 274–282.

33.

Park

and Ham

, Relation network for person re-identification, Proceedings of the AAAI Conference on Artificial Intelligence 34(7) (2020), 11839–11847.

34.

Liu

Zhao

et al., Person re-identification based on multi-scale constraint network, Pattern Recognition Letters 138 (2020), 403–409.

35.

Zhang

Huang

Wei

et al., Appearance feature enhancement for person re-identification, Expert Systems with Applications 163 (2021), 113771.

36.

and Liu

, Guided saliency feature learning for person re-identification in crowded scenes, Computer Vision–ECCV 2020, 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII 16. Springer International Publishing, (2020), 357–373.

37.

Chen

Ding

Xie

Yuan

Chen

Yang

Ren

and Wang

, Abd-net: Attentive but diverse person reidentification, in ICCV, (2019), 8351–8361.

38.

Xia

B.N.

Gong

Zhang

and Poellabauer

, Secondorder non-local attention networks for person re-identification. in ICCV, (2019), 3760–3769.

39.

Zhong

Zheng

Cao

D.L.

et al., Re-ranking person re-identification with k-reciprocal encoding, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Los Alamitos, IEEE Computer Society Press (2017), 3652–3661.

40.

Hanif

Ling

Tian

et al., An Improvement in Multi-query Person Re-identification Using Cumulative Distance Aggregation[C]//. Proceedings of Sixth International Congress on Information and Communication Technology, Springer, Singapore, (2022), 301–309.

41.

Sun

Zhu

and Wang

, Recognizing partial biometric patterns, arXiv preprint arXiv:181007399, (2018).

42.

Liang

and Sun

, Deep spatial feature reconstruction for partial person re-identification: Alignment-free approach, in CVPR, (2018), 7073–7082.

43.

Sun

et al., Perceive where to focus: Learning visibility-aware part-level features for partial person re-identification, in CVPR, (2019), 393–402.

44.

Miao

Liu

et al., Pose-guided feature alignment for occluded person re-identification, Proceedings of the IEEE/CVF International Conference on Computer Vision (2019), 542–551.

45.

Chen

Y.F.

Wang

Sun

X.L.

Fan

and Tang

, Deep attention aware feature learning for person re-identification, arXiv preprint arXiv:200300517, (2020).

46.

S.T.

Luo

Wang

P.C.

Wang

and Jiang

, Transreid: Transformer-based object re-identification, arXiv preprint arXiv:2102.04378, (2021), 2.

47.

Wang

Cheng

et al., RA loss: Relation-aware loss for robust person re-identification, Proceedings of the Asian Conference on Computer Vision (2022), 177–194.

48.

and Zheng

W.S.

, Combined depth space based architecture search for person re-identification, Proceedings of the IEEE, CVF Conference on Computer Vision and Pattern Recognition (2021), 6729–6738.

49.

Zhu

Guo

Zhang

et al., Aaformer: Auto-aligned transformer for person re-identification, IEEE Transactions on Neural Networks and Learning Systems (2023).

50.

Yuan

et al., Denseformer: A dense transformer framework for person re-identification, IET Computer Vision 17(5) (2023), 527–536.

51.

Zhu

Guo

Liu

et al., Identity-guided human semantic parsing for person re-identification, Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16. Springer International Publishing, (2020), 346–363.

Multiple-local feature and attention fused person re-identification method

Abstract

Keywords

1. Introduction

2. Related work

3. Method

4.1 Experimental dataset introduction

Table 1 Parameters of the datasetsThe datasets used in our experiments include market1501, DukeMTMC-reID CUHK03, and MSMT17, which consist of both training and testing sets. Partial-REID and Partial-iLIDS datasets are used for testing occluded scenarios only

Table 2 Effects of CPAA module position on model performance the CPAA module in the 2nd and 3rd layers of the backbone network leads to significant improvement in metrics.adding the CPAA module in the last layer still has a positive effect

Table 4 Performance comparison of different methods on 4 datasetsThe evaluation metrics are shown in bold, while underline indicates the second-best result. ‘–’ denotes that the result is unavailable. Our model achieves competitive performance across the evaluated datasets

Footnotes

Acknowledgments

References

Table 1
Parameters of the datasetsThe datasets used in our experiments include market1501, DukeMTMC-reID CUHK03, and MSMT17, which consist of both training and testing sets. Partial-REID and Partial-iLIDS datasets are used for testing occluded scenarios only

Table 2
Effects of CPAA module position on model performance the CPAA module in the 2nd and 3rd layers of the backbone network leads to significant improvement in metrics.adding the CPAA module in the last layer still has a positive effect

Table 4
Performance comparison of different methods on 4 datasetsThe evaluation metrics are shown in bold, while underline indicates the second-best result. ‘–’ denotes that the result is unavailable. Our model achieves competitive performance across the evaluated datasets