Center Loss Placement Concern

[SuperbTUM]

Hello Hao. I read your journal paper and paid special attention to center loss placement. I saw you decided to put the loss after bn neck to achieve better results, even though this will harm clustering performance. Can you explain a bit more about why this is counter-intuitive as bad clustering for center loss will lead to better re-id accuracy? Is it worth trying stronger inference methods like diminish camera bias, or horizontal flip, to check if the conclusion can be reversed?