Pedestrian Retrieval Using Valuable Absence Augmentation
255
sively complex CNN model is trained, the overfitting might occur, which degenerates the performance of pedestrian retrieval. In order to alleviate the overfitting, the most straightforward method is to increase the number of training
samples. However, it is very expensive to obtain a large number of labeled samples. Data augmentation can extend the training set through transforming the
original training samples, and it is widely used in deep learning, especially in
training deep CNN model. The most common methods of data augmentation
include rotation, adding noises, flipping, cropping, etc. Krizhevsky et al. [11]
randomly flipped samples in the training set to train the deep CNN model. In
order to obtain various training samples, Simonyan et al. [12] randomly cropped
training samples and extracted sub-patch of samples to expand the training set.
In [13], random erasing was proposed as a data augmentation approach, which
randomly selects a rectangle region on the sample and assigns random values.
Xiao et al. [7] combined several databases to jointly train the deep CNN model.
The above-mentioned methods only focus on increasing the training sample number to relieve the overfitting, but it only increases the generalization ability of
the network model to some extent. This is because in the view of the network
model, there is only a subtle difference between original samples and increasing
samples, and that is to say, these increasing samples cannot force the network
model to learn different features.
In this paper, we propose a novel data augmentation method named valuable absence augmentation (VAA) to increase the number of training samples
and meanwhile utilize these increasing samples to train the network in order to
force the network to learn variable features. Specifically, we first train a base
network model using the original training set. Next, we obtain the attention
map extracted from convolutional activation maps of the trained base network
model. Afterward, we take valuable parts of the attention map and project these
parts to the original sample and set the corresponding position of the original
sample to zero. As a result, we can obtain a large number of images without
valuable parts for the base network model. Finally, we combine these samples
with original samples to fine-tune the base network model so that the network
model can learn new discriminative features from valuable absence parts of the
sample. Hence, the generalization ability of the network model can be improved.
We evaluate the proposed VAA on a large-scale pedestrian retrieval database
(Market-1501), and extensive experiments show that the proposed VAA is effective for reducing the risk of overfitting and improving the generalization ability
of the network model. In addition, during the test stage, we also evaluate the
impact of removing valuable parts with different proportions from the original
sample. Experimental results show that the performance of pedestrian retrieval
decreases with the increase of removing valuable parts ratio.
In summary, the contribution of this paper is: (1) We propose VAA to increase
the number of training set in order to mitigate the risk of overfitting; (2) we utilize these increasing samples and original samples to fine-tune the base network,
which can improve the generalization ability of the network model; and (3) we
prove the effectiveness of VAA on a large-scale database.
255
sively complex CNN model is trained, the overfitting might occur, which degenerates the performance of pedestrian retrieval. In order to alleviate the overfitting, the most straightforward method is to increase the number of training
samples. However, it is very expensive to obtain a large number of labeled samples. Data augmentation can extend the training set through transforming the
original training samples, and it is widely used in deep learning, especially in
training deep CNN model. The most common methods of data augmentation
include rotation, adding noises, flipping, cropping, etc. Krizhevsky et al. [11]
randomly flipped samples in the training set to train the deep CNN model. In
order to obtain various training samples, Simonyan et al. [12] randomly cropped
training samples and extracted sub-patch of samples to expand the training set.
In [13], random erasing was proposed as a data augmentation approach, which
randomly selects a rectangle region on the sample and assigns random values.
Xiao et al. [7] combined several databases to jointly train the deep CNN model.
The above-mentioned methods only focus on increasing the training sample number to relieve the overfitting, but it only increases the generalization ability of
the network model to some extent. This is because in the view of the network
model, there is only a subtle difference between original samples and increasing
samples, and that is to say, these increasing samples cannot force the network
model to learn different features.
In this paper, we propose a novel data augmentation method named valuable absence augmentation (VAA) to increase the number of training samples
and meanwhile utilize these increasing samples to train the network in order to
force the network to learn variable features. Specifically, we first train a base
network model using the original training set. Next, we obtain the attention
map extracted from convolutional activation maps of the trained base network
model. Afterward, we take valuable parts of the attention map and project these
parts to the original sample and set the corresponding position of the original
sample to zero. As a result, we can obtain a large number of images without
valuable parts for the base network model. Finally, we combine these samples
with original samples to fine-tune the base network model so that the network
model can learn new discriminative features from valuable absence parts of the
sample. Hence, the generalization ability of the network model can be improved.
We evaluate the proposed VAA on a large-scale pedestrian retrieval database
(Market-1501), and extensive experiments show that the proposed VAA is effective for reducing the risk of overfitting and improving the generalization ability
of the network model. In addition, during the test stage, we also evaluate the
impact of removing valuable parts with different proportions from the original
sample. Experimental results show that the performance of pedestrian retrieval
decreases with the increase of removing valuable parts ratio.
In summary, the contribution of this paper is: (1) We propose VAA to increase
the number of training set in order to mitigate the risk of overfitting; (2) we utilize these increasing samples and original samples to fine-tune the base network,
which can improve the generalization ability of the network model; and (3) we
prove the effectiveness of VAA on a large-scale database.