Details for Person Re-identification
Baseline
Zhong Zhang
1,2(B) , Haijia Zhang
1,2 , and Shuang Liu
1,2
1 Tianjin Key Laboratory of Wireless Mobile Communications
and Power Transmission, Tianjin Normal University, Tianjin, China
haijia27zhang@gmail.com, shuangliu.tjnu@gmail.com
2 College of Electronic and Communication Engineering,
Tianjin Normal University, Tianjin, China
zhong.zhang8848@gmail.com
Abstract. In this paper, we evaluate the performance of person reidentification (Re-ID) baseline under different implementation details
including resize ratio of the pedestrian image, batch size and basic learning rate. To this end, we employ ResNet-50 as the classification model by
modifying the original FC layer and apply a classifier to compute identity prediction values. We perform amounts of experiments to assess the
effects of these implementation details on Market-1501 and experimental results show that these implementation details are very important for
person Re-ID.
Keywords: Person re-identification · Baseline · Implementation
details
1 Introduction
Person re-identification (Re-ID) targets to spot and return an image of pedestrian with the same identity and different camera view from the gallery set. Due
to complex variations of pedestrian images in pose, camera conditions and background, hence, person Re-ID is a challenging task. It has important applications
in surveillance system [1] and human activity analysis [2], and therefore, both
academic and industry pay more and more attention on person Re-ID. Before
researchers of the image classification and object recognition fields turn to convolutional neural network (CNN) and deep learning, handcrafted features are used
to perform the feature representation [3–6]. Many approaches are proposed for
person Re-ID [7–11]. Recently, CNN and deep learning have been widely applied
in person Re-ID task, and some large-scale person Re-ID datasets including
Market-1501 [12], CUHK03 [13] and DukeMTMC-reID [14] are issued to train
deep model. Deep features learned from CNN are robust to variances of pedestrian images, and therefore, CNN-based methods have dominated person Re-ID
c
Springer Nature Singapore Pte Ltd. 2020
Q. Liang et al. (Eds.): Artificial Intelligence in China, LNEE 572, pp. 452–457, 2020.
https://doi.org/10.1007/978-981-15-0187-6_54
Baseline
Zhong Zhang
1,2(B) , Haijia Zhang
1,2 , and Shuang Liu
1,2
1 Tianjin Key Laboratory of Wireless Mobile Communications
and Power Transmission, Tianjin Normal University, Tianjin, China
haijia27zhang@gmail.com, shuangliu.tjnu@gmail.com
2 College of Electronic and Communication Engineering,
Tianjin Normal University, Tianjin, China
zhong.zhang8848@gmail.com
Abstract. In this paper, we evaluate the performance of person reidentification (Re-ID) baseline under different implementation details
including resize ratio of the pedestrian image, batch size and basic learning rate. To this end, we employ ResNet-50 as the classification model by
modifying the original FC layer and apply a classifier to compute identity prediction values. We perform amounts of experiments to assess the
effects of these implementation details on Market-1501 and experimental results show that these implementation details are very important for
person Re-ID.
Keywords: Person re-identification · Baseline · Implementation
details
1 Introduction
Person re-identification (Re-ID) targets to spot and return an image of pedestrian with the same identity and different camera view from the gallery set. Due
to complex variations of pedestrian images in pose, camera conditions and background, hence, person Re-ID is a challenging task. It has important applications
in surveillance system [1] and human activity analysis [2], and therefore, both
academic and industry pay more and more attention on person Re-ID. Before
researchers of the image classification and object recognition fields turn to convolutional neural network (CNN) and deep learning, handcrafted features are used
to perform the feature representation [3–6]. Many approaches are proposed for
person Re-ID [7–11]. Recently, CNN and deep learning have been widely applied
in person Re-ID task, and some large-scale person Re-ID datasets including
Market-1501 [12], CUHK03 [13] and DukeMTMC-reID [14] are issued to train
deep model. Deep features learned from CNN are robust to variances of pedestrian images, and therefore, CNN-based methods have dominated person Re-ID
c
Springer Nature Singapore Pte Ltd. 2020
Q. Liang et al. (Eds.): Artificial Intelligence in China, LNEE 572, pp. 452–457, 2020.
https://doi.org/10.1007/978-981-15-0187-6_54