Furthermore, a novel Intra-class Hard Example Mining (IHEM) loss function is proposed, which penalizes the cosine distance between the hard examples and their class centers to enhance the ...