An Uncertainly Dynamic Loss Correction and Global Sample Selection Method for Webly Supervised Fine-Grained Visual Classification.

To distinguish the subtle differences among fine-grained categories, a large amount of well-labeled images are typically required. However, manual annotations for fine-grained categories is an extremely difficult task as it usually has a high demand for professional knowledge. To this end, we propose to directly leverage web images for fine-grained visual recognition. Nevertheless, directly utilizing web images for training fine-grained classification models tends to have poor performance due to the existence of label noise. In this work, we propose an end-to-end method by combining uncertainly dynamic loss correction and global sample selection to solve the problem of label noise. Specifically, we leverage a deep neural network to predict all samples, record the predictions of several recent epochs and calculate the uncertainly dynamic loss for global sample selection in the whole epoch. We conduct experiments on three commonly used noisy fine-grained datasets Web-Aircraft, Web-Bird and Web-Cars. The average classification accuracy is 75.40%, 78.53% and 82.19%, which has 1.20%, 2.16% and 3.43% improvements, respectively. [ABSTRACT FROM AUTHOR]