258
X. Hao et al.
3 Experiments
In this section, we conduct extensive experiments on a large-scale pedestrian
retrieval database (Market-1501 [15]). We first evaluate the influence of the
pedestrian valuable parts for network performance. Then, we evaluate the proposed VAA on this database. During the evaluation process, the cumulative
match characteristic and the mean average precision (mAP) are treated as
criteria.
3.1 Market-1501 Database
Market-1501 is a large-scale pedestrian retrieval database. It is collected from
six different cameras and contains 32,668 pedestrian images with 1501 identities.
This database has 12,936 pedestrian images of 751 pedestrian in the training
set, 19,732 pedestrian images of 750 pedestrian in the testing set, and 3368
pedestrian images as the query set. Each pedestrian has an average of 17.2
images, and each pedestrian image is detected by the Deformable Parts Model
(DPM) detector [16], which is close to the real environment. We report the
Rank-1 accuracy and mAP for the sake of fair comparison with other methods.
3.2 Implementation Details
When training the base network model, we utilize random horizontal flipping as a
data augmentation method and resize the size of pedestrian image to 256×128 as
the input of the base network model. We set the batch size to 32 and the number
of epoch to 60 and use stochastic gradient descent (SGD) to optimize this model’s
parameters. As for the learning rate, we set the last layer (FC) to 0.1 and other
layers to 0.01, and the learning rate gradually decreases at the rate of 0.1 after
40 epochs. In the second phase, we select convolutional activation maps from the
output of Conv 2 and utilize Eq. 1 to obtain the attention map. We project the
first 5% valuable parts of the attention map to original pedestrian image and set
corresponding positions to zero. As a result, we obtain new samples. In the last
phase, all parameter settings are the same as in the first phase. During testing,
we extract the 2048-dim feature vector from the output of pooling5 to represent
pedestrians and compute the similarity of features by the Euclidean distance.
3.3 Evaluation of Valuable Parts for Network Performance
We evaluate the influence of the pedestrian valuable parts for the network performance in this subsection, and experiment results are listed in Table 1. We
can see that the performance of pedestrian retrieval becomes worse and worse
as the proportion of removing valuable parts increases. Specifically, when 5%
valuable parts are removed, the performance of pedestrian retrieval decreases by
only 7% in the Rank-1 accuracy, but when 20% valuable parts are removed, the
performance of pedestrian retrieval drops drastically and only reaches the result
of 52.2% Rank-1 accuracy. This indicates that the valuable part has a crucial
influence on the performance of the network model.
X. Hao et al.
3 Experiments
In this section, we conduct extensive experiments on a large-scale pedestrian
retrieval database (Market-1501 [15]). We first evaluate the influence of the
pedestrian valuable parts for network performance. Then, we evaluate the proposed VAA on this database. During the evaluation process, the cumulative
match characteristic and the mean average precision (mAP) are treated as
criteria.
3.1 Market-1501 Database
Market-1501 is a large-scale pedestrian retrieval database. It is collected from
six different cameras and contains 32,668 pedestrian images with 1501 identities.
This database has 12,936 pedestrian images of 751 pedestrian in the training
set, 19,732 pedestrian images of 750 pedestrian in the testing set, and 3368
pedestrian images as the query set. Each pedestrian has an average of 17.2
images, and each pedestrian image is detected by the Deformable Parts Model
(DPM) detector [16], which is close to the real environment. We report the
Rank-1 accuracy and mAP for the sake of fair comparison with other methods.
3.2 Implementation Details
When training the base network model, we utilize random horizontal flipping as a
data augmentation method and resize the size of pedestrian image to 256×128 as
the input of the base network model. We set the batch size to 32 and the number
of epoch to 60 and use stochastic gradient descent (SGD) to optimize this model’s
parameters. As for the learning rate, we set the last layer (FC) to 0.1 and other
layers to 0.01, and the learning rate gradually decreases at the rate of 0.1 after
40 epochs. In the second phase, we select convolutional activation maps from the
output of Conv 2 and utilize Eq. 1 to obtain the attention map. We project the
first 5% valuable parts of the attention map to original pedestrian image and set
corresponding positions to zero. As a result, we obtain new samples. In the last
phase, all parameter settings are the same as in the first phase. During testing,
we extract the 2048-dim feature vector from the output of pooling5 to represent
pedestrians and compute the similarity of features by the Euclidean distance.
3.3 Evaluation of Valuable Parts for Network Performance
We evaluate the influence of the pedestrian valuable parts for the network performance in this subsection, and experiment results are listed in Table 1. We
can see that the performance of pedestrian retrieval becomes worse and worse
as the proportion of removing valuable parts increases. Specifically, when 5%
valuable parts are removed, the performance of pedestrian retrieval decreases by
only 7% in the Rank-1 accuracy, but when 20% valuable parts are removed, the
performance of pedestrian retrieval drops drastically and only reaches the result
of 52.2% Rank-1 accuracy. This indicates that the valuable part has a crucial
influence on the performance of the network model.