MobileNets: Efficient Neural Networks for Mobile Vision

We present MobileNets, a class of efficient models designed for mobile and embedded vision applications, utilizing a streamlined architecture with depth-wise separable convolutions to create lightweight deep neural networks. Two global hyper-parameters are introduced to balance latency and accuracy, enabling model builders to select the appropriate model size based on specific constraints. Extensive experiments demonstrate strong performance on ImageNet classification and highlight the versatility of MobileNets across various applications, including object detection, fine-grain classification, face attributes, and large-scale geo-localization.