TriMoFaceNet: a trimmed MobileFaceNet architecture for resource-constrained environments

As one of the most popular vision applications, face verification systems can be needed in all kinds of scenarios, including resource-constrained environments. In this paper, we propose a trimmed variant of MobileFaceNet, which can perform as well as MobileFaceNet but with fewer memory and computational requirements. We essentially make two changes to the bottlenecks: reversing the expansion-projection process and adding one more depth-wise convolution layer. While the first one significantly reduces the number of parameters and calculations, the second one increases them. We try to strike a balance between the two to have a complex enough network architecture that can perform face verification as well as the original one. We could cut down the model size by almost a quarter in this manner, and we name the resultant network TriMoFaceNet.