SSCRL: fine-grained object retrieval with switched shifted centralized ranking loss

Image retrieval is an attractive task in computer vision that aims at browsing, searching, and returning images from a large database of digital images after delivering a retrieval query. Numerous works have focused on fine-grained object retrieval (FGOR) because it is extremely challenging and of great value in practical application. Due to the large diversity within a class and the small diversity across different classes of fine-grained objects data, a convolutional neural network (CNN) is a powerful extractor that can be used to obtain fine-grained features for distinguishing tiny variations between classes. As an indispensable part of a convolutional neural network model, the loss function is of critical importance for feature extraction. In this work, based on the global structure loss function, we propose a variant of softmax loss, named switched shifted softmax loss, to potentially reduce the overfitting phenomenon of the model. Comparative experiments with different backbone structures verify that the developed loss function with trivial transformation enhances the fine-grained retrieval performance of deep learning methods 1 . Furthermore, additional experiments of fine-grained object classification and person re-identification (re-ID) prove that our method has a wide spectrum of applicability to other tasks.