Gaussian-based probability fusion for person re-identification with Taylor angular margin loss

Similar Image Retrieval Algorithm Based on Feature Fusion and Locality-Sensitive Hash

“…Due to the existence of similar images in real scenes with large image background differences and high background similarity of heterogeneous images, which makes the inter-class distance of heterogeneous images small, and the standard cross-entropy loss function does not constrain the inter-class distance, the extracted features are easy to be confused between classes. Therefore, this paper uses Arcface loss [10] as the loss function, which can achieve good results under weak labels [11], increases the inter-class distance and further converges the intra-class distance by adding an angle penalty term for the dissimilar classes in the angle domain [12]. The Arcface loss used for the loss function of the feature extraction network in this paper is as follows: The class activation map [13] is a weighted linear sum of visual patterns present at different spatial locations that can suggest which region of the image our view neural network is focusing on, and the feature heat map of the features output by the two-branch convolutional layer is shown in Figure 5, where it can be visualized that the two features can play a complementary role in the characterization of the image content.…”