Motivation

Traditional computer vision techniques were used to perform most computer vision tasks, such as object detection and segmentation. The performance of these traditional computer vision techniques was good but it was never close to being usable in real time, for example by autonomous cars. In 2012, Alex Krizhevsky introduced CNNs, which made a breakthrough on the ImageNet competition by enhancing the object classification error from 26% to 15%. CNNs have been widely used since then and different variations have been discovered. It has even outperformed the human classification error over the ImageNet competition, as shown in the following diagram:

Figure 9.4: Classification error over time with human level error marked in red

Applications of CNNs

Since the breakthrough the CNNs achieved in different domains of computer vision and even natural language processing, most companies have integrated this deep learning solution into their computer vision echo system. For example, Google...