Circle-Based Ratio Loss for Person Reidentification

Figure 4

The learning rate curve with the warm-up strategy. During the first 20 epochs, the learning rate linearly increases from a small value of 10⁻⁵ to 10⁻³, and it remains unchanged in the latter 70 epochs. Then the learning rate is decayed by 0.1 at 90^th and 130^th, respectively, to fine-tune the network parameters.