Evolution path to MobileNets

CNNs present a promising future for computer vision. CNNs have laid out a benchmark for complex computer vision tasks such as detection and recognition with their remarkable performance in the ILSVRC competition over consecutive years. But the computation power required by these CNN models has always been quite high. This could lead to a major setback for the commercial use of CNNs. Almost all object detection-related tasks in the real world are performed through portable devices, such as mobile phones, surveillance cameras, or any other embedded device. These devices have limited computational abilities and memory. To make any deep learning network running on a portable device, the network weights and the number of calculations occurring in the network (that is, the number of parameters in the network) have to very small. CNNs have millions of...