In the previous section, we saw that two-stage detectors suffer from the issue of slower prediction time and harder training by splitting the network into two. In recently proposed networks like Single Shot Multibox Detectors (SSD)[3], the prediction time is reduced by removing the intermediate stage and the training is always end-to-end. These networks have shown effectiveness by running on smartphones as well as low-end computation units:
An abstract view of the network is shown in the preceding figure. The overall output of the network is same as two-stage, the class probability for the object and bounding box coordinates ...
