The makers of MobileNetV2 also made real-time object detection possible for mobile devices. They introduced a combination of the SSD Object Detector and MobileNetV2, which is called SSDLite. Remember that in Chapter 4, CNN Architecture, we used ssd_mobilenetv2 for object detection. It is the same as SSDLite. The reason for choosing SSD is quite simple. SSD is built independent of the base network and hence the convolutions are replaced by depth-wise separable convolution. The first layer of SSDLite is attached to the expansion of layer 15 of MobileNetV2. Replacing standard convolutions with depth-wise separable convolution significantly reduces the number of parameters that are required by the network for object detection.
The following table shows a comparison of the number of parameters and multiplication operations required by the original SSD network...
