4.2 RetinaNet model
Since the one-stage detector is not as accurate as the two-stage detector, T. -Y. Lin et al. [22] proposed a new loss function: focal loss and designed a dense detector: RetinaNet. By reducing the weight of the easily classified samples to pay more attention to the difficult-to-classify samples, the RetinaNet can achieve not only the speed of the first-level detector, but also the accuracy of the second-level detector. RetinaNet is essentially composed of resnet residual network, FPN with two FCN sub-networks. The structure of RetinaNet model is shown in Figure 7.