focal loss fix single-shot detection
RetinaNet: how Focal Loss fixes Single-Shot Detection
Neural networks can be used to solve classification problems (predict classes) and regression problems (predict continuous values). Today we will be doing both at the same time. We start with a simplified task: detect and classify one single object in an image instead of several objects. How does an object detection dataset look like? Well, the inputs to our model are of course images and the labels are typically four values that describe a ground truth bounding box plus a category the object in this box belongs two.