| Autor: |
Guan-Wen Zhang, Jien Kato, Yu Wang, Kenji Mase |
| Předmět: |
|
| Zdroj: |
Computer Systems Science & Engineering; Jul2014, Vol. 29 Issue 4, p265-274, 10p |
| Abstrakt: |
In this paper, we propose a deep learning based method for people re-identification tasks. We design a deep convolutional neural network (CNN) which integrates both feature learning and re-identification in one unified framework. In order to make the outputs of the CNN compensate with re-identification tasks, we introduce a linear support vector machine to replace the softmax layer in traditional approaches. This makes the outputs can be used to measure the similarity between images. Besides, to robustly learn a large number of parameters for the network from a small set of training data, we introduced three training strategies: unsupervised pre-training, dropout and data augmentation. These strategies efficiently prevent the network training from overfitting. The proposed deep CNN was evaluated on challenging VIPeR dataset and CAVIAR4REID dataset. The experiment results confirm the superior of the proposed method. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Supplemental Index |
| Externí odkaz: |
|
