Görünür ve Kızılötesi Görüntülerde Kişiyi Yeniden Tanıma
Özet
Person re-identification and cross-modality person re-identification are computer vision tasks that aim to accurately match images of individuals. Visible-infrared cross-modality person re-identification is a more challenging problem compared to person re-identification due to the absence of color information and the differences in modalities. With the emergence of deep learning-based approaches, rapid progress has been made in the field of cross-modality person re-identification in recent years.
Within the scope of this thesis, a layer is proposed that performs person identification using distance metrics on prototypes. The performance of the proposed layer is evaluated with various distance metrics and update methods. Alignment and attention mechanisms are investigated, and the effectiveness of these structures is evaluated. An adaptive weighting scheme is proposed for the horizontal part splitting approach, aiming to focus the deep neural network on local features. The effect on performance of applying loss functions on low-level and mid-level features of deep neural networks is investigated. Furthermore, a data augmentation method called horizontal stripe augmentation is proposed. This method replaces horizontal parts of an image with corresponding cross-modality parts of the same individual. With the proposed data augmentation method, the neural network is encouraged to focus more on local features, and the modality gap is alleviated. The proposed method outperforms other cross-modality data augmentation methods used in the literature. Lastly, an average precision-based loss function is employed for training a deep neural network. Margin terms that make ranking difficult for positive and cross-modality samples are introduced into the loss function, which includes approximated average precision. The margin-enhanced approximated average precision increases the separation of hard samples without the need for additional distance-based loss function. The performance of the proposed method is evaluated with different margin values and hyperparameter settings. Experimental results demonstrate the effectiveness of the new margin-enhanced loss function.
