Directional statistics-based deep metric learning for image classification and retrieval

L2-normalization is an effective method to enhance the discriminant power of deep representation learning. However, without exploiting the geometric properties of the feature space, the generally used gradient based optimization methods are failed to track the global information during training. In this paper, we propose a novel deep metric learning model based on the directional distribution. By defining the loss function based on the von Mises–Fisher distribution, we propose an effective alternative learning algorithm by periodically updating the class centers. The proposed metric learning not only captures the global information about the embedding space but also yields an approximate representation of the class distribution during training. Considering classification and retrieval tasks, our experiments on benchmark datasets demonstrate an improvement from the proposed algorithm. Particularly, with a small number of convolutional layers, a significant accuracy upsurge can be observed compared to widely used gradient-based methods.