MobileNetV2: Inverted Residuals and Linear Bottlenecks

By Mark Sandler et al

Read the original document by opening this link in a new tab.

Table of Contents

Abstract
1. Introduction
2. Related Work
3. Preliminaries, discussion and intuition
4. Model Architecture
5. Implementation Notes

Summary

MobileNetV2: Inverted Residuals and Linear Bottlenecks presents a new mobile architecture that enhances the performance of mobile models across various tasks and benchmarks. The paper introduces MobileNetV2, a novel framework called SSDLite, and Mobile DeepLabv3 for object detection and semantic segmentation models. The architecture is based on inverted residual structures with linear bottlenecks, emphasizing the importance of non-linear removal in narrow layers. By decoupling input/output domains from the transformation's expressiveness, the paper achieves improved performance on ImageNet classification, COCO object detection, and VOC image segmentation. The document discusses the depthwise separable convolutions, linear bottlenecks, and inverted residuals, providing insights into optimizing neural architectures for mobile and resource-constrained environments.