Details for Person Re-identification Baseline
453
community [15–18]. Specifically, these approaches are based on the deep classification model which is treated as the baseline.
In this paper, we perform amounts of experiments to assess influence of different implementation details on the baseline including resize ratio of the pedestrian
image, batch size and basic learning rate. We make two contributions as follows:
• We detail the architecture of the baseline to obtain a clear understanding of
the classification model.
• We verify various implementation details could have some effects on the performance of the baseline.
We divide the rest of paper into three sections. Firstly, we detail the framework of the baseline in Sect. 2. Afterwards, experimental results conducted on
Market-1501 is provided in Sect. 3. Finally, we sum up our work in Sect. 4.
2 Approach
In this section, we utilize the ResNet-50 [19] as the baseline because of its excellent characteristics as well as relatively succinct framework.
2.1 ResNet
The residual network (ResNet) is proposed by He et al. [19], and they participate
in the ILSVRC with it. The performance of ResNet in image classification, target
detection and other tasks greatly surpasses the performance level of the previous
year’s competition. The obvious feature of ResNet is that it has a fairly deep
depth, from 34 to 152 levels. The depth of ResNet is amazing, and the deep
depth makes the network very powerful. The ResNet has several characteristics
as follows:
• Firstly, it has compact structure without increasing extra parameters.
• Secondly, it increases the number of feature maps layer by layer to retain more
useful information.
• Thirdly, BN and global average pooling layers are utilized for regularization
to accelerate model convergence.
2.2 From ResNet-50 to the Baseline
The 50-layer ResNet (ResNet-50) has the suitable network architecture for person Re-ID, and therefore, many approaches [20–22] employ ResNet-50 as backbone to modify and match the person Re-ID task. We just make one modification
on the ResNet-50 to match the classification task. Concretely, the number of neurons in the original fully connected (FC) layer of ResNet-50 is changed from 1000
to 512. Hence, we could obtain a feature vector f with the dimension of 512 from
the modified ResNet-50. Afterwards, we design a classifier which is composed of
one FC layer and the softmax function to compute identity prediction values.
453
community [15–18]. Specifically, these approaches are based on the deep classification model which is treated as the baseline.
In this paper, we perform amounts of experiments to assess influence of different implementation details on the baseline including resize ratio of the pedestrian
image, batch size and basic learning rate. We make two contributions as follows:
• We detail the architecture of the baseline to obtain a clear understanding of
the classification model.
• We verify various implementation details could have some effects on the performance of the baseline.
We divide the rest of paper into three sections. Firstly, we detail the framework of the baseline in Sect. 2. Afterwards, experimental results conducted on
Market-1501 is provided in Sect. 3. Finally, we sum up our work in Sect. 4.
2 Approach
In this section, we utilize the ResNet-50 [19] as the baseline because of its excellent characteristics as well as relatively succinct framework.
2.1 ResNet
The residual network (ResNet) is proposed by He et al. [19], and they participate
in the ILSVRC with it. The performance of ResNet in image classification, target
detection and other tasks greatly surpasses the performance level of the previous
year’s competition. The obvious feature of ResNet is that it has a fairly deep
depth, from 34 to 152 levels. The depth of ResNet is amazing, and the deep
depth makes the network very powerful. The ResNet has several characteristics
as follows:
• Firstly, it has compact structure without increasing extra parameters.
• Secondly, it increases the number of feature maps layer by layer to retain more
useful information.
• Thirdly, BN and global average pooling layers are utilized for regularization
to accelerate model convergence.
2.2 From ResNet-50 to the Baseline
The 50-layer ResNet (ResNet-50) has the suitable network architecture for person Re-ID, and therefore, many approaches [20–22] employ ResNet-50 as backbone to modify and match the person Re-ID task. We just make one modification
on the ResNet-50 to match the classification task. Concretely, the number of neurons in the original fully connected (FC) layer of ResNet-50 is changed from 1000
to 512. Hence, we could obtain a feature vector f with the dimension of 512 from
the modified ResNet-50. Afterwards, we design a classifier which is composed of
one FC layer and the softmax function to compute identity prediction values.