MobileNetV2 is very similar to the original MobileNet, except that it uses inverted residual blocks with bottlenecking features. It has a drastically lower parameter count than the original MobileNet. MobileNets support any input size greater than 32 x 32, with larger image sizes offering better performance. MobileNetV2 uses both depthwise and pointwise convolutions to produce depthwise-separable convolutions which allows the model to notice more features of our data.
The MobileNetV2 research was carried out by Mark Sandler and Andrew Howard which by the way led to a great improvement and efficiency of the model compared to the MobileNetV1. The major improvements was the residual skip connection which allows our model not to lose vital information by passing information from the input via the
residual connection directly to the output just like Resnet. They also added bottlenecking features which was an improvement from just the depthwise convolutions of the MobileNetV1, the bottlenecking features consist of a depthwise convolutions and pointwise convolutions which reduces the size and increase the size of the data respectively.
