This paper introduces MobileNetV2, a new mobile architecture that improves performance on multiple tasks and benchmarks across a range of model sizes. The architecture is based on inverted residuals with linear bottlenecks, where shortcut connections are between thin bottleneck layers. Intermediate expansion layers use lightweight depthwise convolutions to filter features, and non-linearities are removed in the narrow layers to maintain representational power. This design allows decoupling of input/output domains from the expressiveness of the transformation, facilitating further analysis. The paper also discusses efficient applications of these models to object detection (SSDLite) and semantic segmentation (Mobile DeepLabv3). Experiments show that MobileNetV2 outperforms state-of-the-art models in terms of accuracy, computational cost, and parameter count, making it suitable for mobile and resource-constrained environments.This paper introduces MobileNetV2, a new mobile architecture that improves performance on multiple tasks and benchmarks across a range of model sizes. The architecture is based on inverted residuals with linear bottlenecks, where shortcut connections are between thin bottleneck layers. Intermediate expansion layers use lightweight depthwise convolutions to filter features, and non-linearities are removed in the narrow layers to maintain representational power. This design allows decoupling of input/output domains from the expressiveness of the transformation, facilitating further analysis. The paper also discusses efficient applications of these models to object detection (SSDLite) and semantic segmentation (Mobile DeepLabv3). Experiments show that MobileNetV2 outperforms state-of-the-art models in terms of accuracy, computational cost, and parameter count, making it suitable for mobile and resource-constrained environments.