RMNet is custom backbone designed for the fast but still accurate inference. The RMNet architecture is inspired by ResNet and MobileNet architectures. From the first one we get the idea to use residual blocks. In our case the residual block consists of the sequence of convolutions with spatial resolutions 1x1, 3x3, 1x1, Elu activation function and skip connection. To speed up the computations we also use the depth-wise convolutions like in MobileNet instead of default 3x3 conv. Finally, the maximal number of channels is limited to 256. It allows us to increase the depth of network to be close to ResNet101 by the receptive filed.
